September 14, 2024

How Much GPU Memory is Required to Run a Large Language Model?

With the growing importance of LLMs in AI-driven applications, developers and companies are deploying models like GPT-4, LLaMA, and OPT-175B in real-world scenarios. However, one of the most overlooked aspects of deploying these models is understanding how much GPU memory is needed to serve them effectively. Miscalculating memory requirements can cost you significantly more in hardware or cause downtime due to insufficient resources.

In this article, we’ll explore the key components contributing to GPU memory usage during LLM inference and how you can accurately estimate your GPU memory requirements. We’ll also discuss advanced techniques to reduce memory wastage and optimize performance. Let’s dive in!

Understanding GPU Memory Requirements for LLMs

LLMs rely heavily on GPU resources for inference. GPU memory consumption for serving LLMs can be broken down into four key components:

Model Parameters (Weights)

Key-Value (KV) Cache Memory

Activations and Temporary Buffers

Memory Overheads

Let’s examine each of these in more detail and see how they contribute to the total memory footprint.

Model Parameters (Weights)

Model parameters are the neural network’s learned weights. These weights are stored in GPU memory during inference, and their size is directly proportional to the number of parameters in the model.

How Model Size Impacts Memory

A typical inference setup uses each parameter’s FP16 (half-precision) format to save memory while maintaining acceptable precision. Each parameter requires 2 bytes in FP16 format.

For example:

A small LLM with 345 million parameters would require:

345 million × 2 bytes = 690 MB of GPU memory.

A larger model like LLaMA 13B (13 billion parameters) would require:

13 billion × 2 bytes = 26 GB of GPU memory.

For massive models like GPT-3, which has 175 billion parameters, the memory requirement becomes:

175 billion × 2 bytes = 350 GB.

Clearly, larger models demand significantly more memory, and distributing the model across multiple GPUs becomes necessary for serving these larger models.

Key-Value (KV) Cache Memory

The KV cache stores the intermediate key and value vectors generated during the model’s inference process. This is essential for maintaining the context of the sequence being generated. As the model generates new tokens, the KV cache stores previous tokens, allowing the model to reference them without re-calculating their representations.

How Sequence Length and Concurrent Requests Impact KV Cache

Sequence Length: Longer sequences require more tokens, leading to a larger KV cache.

Concurrent Users: Multiple users increase the number of generated sequences, which multiplies the required KV cache memory.

Calculating KV Cache Memory

Here’s a simplified way to calculate the KV cache memory:

For each token, a key and value vector are stored.

The number of vectors per token is equal to the number of layers in the model (L), and the size of each vector is the hidden size (H).

For example, consider a LLaMA 13B model with:

L = 40 layers

H = 5120 dimensions

The KV cache per token is calculated as:

Key Vector: 40 × 5120 = 204,800 elements

FP16 requires 204,800 × 2 bytes = 400 KB per key vector.

The value vector needs the same memory, so the total KV cache memory per token is 800 KB.

For a sequence of 2000 tokens:

2000 tokens × 800 KB = 1.6 GB per sequence.

If the system serves 10 concurrent users, the total KV cache memory becomes:

1.6 GB × 10 = 16 GB of GPU memory for KV cache alone.

Activations and Temporary Buffers

Activations are the outputs of the neural network layers during inference. Temporary buffers store intermediate results during matrix multiplications and other computations.

While activations and buffers usually consume less memory than model weights and KV cache, they still account for approximately 5-10% of the total memory.

Memory Overheads and Fragmentation

Memory overheads come from how memory is allocated. Fragmentation can occur when memory blocks are not fully utilized, leaving gaps that cannot be used efficiently.

Internal Fragmentation: This occurs when memory blocks are not filled.

External Fragmentation: This happens when free memory is split into non-contiguous blocks, making it difficult to allocate large chunks of memory when needed.

Inefficient memory allocation can waste 20-30% of total memory, reducing performance and limiting scalability.

Calculating Total GPU Memory

Now that we understand the components, we can calculate the total GPU memory required for serving an LLM.

For example, let’s calculate the total memory needed for a LLaMA 13B model with the following assumptions:

The total memory required would be:

26 GB + 16 GB + 9.2 GB (for activations and overheads) = 101.2 GB.

Thus, under this scenario, you would need at least 3 A100 GPUs (each with 40 GB of memory) to serve an LLaMA 13B model.

Challenges in GPU Memory Optimization

Over-allocating memory for the key-value (KV) cache, or experiencing fragmentation within the memory, can significantly reduce the capacity of a system to handle a large number of requests. These issues often arise in systems dealing with complex tasks, especially in natural language processing (NLP) models or other AI-based frameworks that rely on efficient memory management. Furthermore, when advanced decoding algorithms, such as beam search or parallel sampling, are used, the memory demands grow exponentially. This is because each sequence being processed requires a dedicated KV cache, resulting in even greater pressure on the system’s memory resources. Consequently, both over-allocation and fragmentation can lead to performance bottlenecks, restricting scalability and reducing efficiency.

Memory Optimization Techniques

PagedAttention: Reducing Memory Fragmentation with Paging

PagedAttention is a sophisticated memory management technique inspired by how operating systems handle virtual memory. When we think of computer memory, it’s easy to imagine it as one big block where data is stored in a neat, continuous fashion. However, when dealing with large-scale tasks, especially in machine learning or AI models, allocating such large chunks of memory can be inefficient and lead to memory fragmentation.

What is Memory Fragmentation?

Fragmentation happens when memory is allocated in a way that leaves small, unusable gaps between different data blocks. Over time, these gaps can build up, making it harder for the system to find large, continuous memory spaces for new data. This leads to inefficient memory use and can slow down the system, limiting its ability to process large numbers of requests or handle complex tasks.

How Does PagedAttention Work?

PagedAttention solves this by breaking down the key-value (KV) cache—used for storing intermediate information in attention mechanisms—into smaller, non-contiguous blocks of memory. Rather than requiring one large, continuous block of memory, it pages the cache, similar to how an operating system uses virtual memory to manage data in pages.

Dynamically Allocated: The KV cache is broken into smaller pieces that can be spread across different parts of memory, making better use of available space.

Reduced Fragmentation: By using smaller blocks, it reduces the number of memory gaps, leading to better memory utilization. This helps prevent fragmentation, as there’s no need to find large, continuous blocks of memory for new tasks.

Improved Performance: Since memory is allocated more efficiently, the system can handle more requests simultaneously without running into memory bottlenecks.

vLLM: A Near-Zero Memory Waste Solution

Building on the concept of PagedAttention, vLLM is a more advanced technique designed to optimize GPU memory usage even further. Modern machine learning models, especially those that run on GPUs (Graphics Processing Units), are incredibly memory-intensive. Inefficient memory allocation can quickly become a bottleneck, limiting the number of requests a system can process or the size of batches it can handle.

What Does vLLM Do?

vLLM is designed to minimize memory waste to nearly zero, allowing systems to handle more data, larger batches, and more requests with fewer resources. It achieves this by making memory allocation more flexible and reducing the amount of memory that goes unused during processing.

Key Features of vLLM:

Dynamic Memory Allocation:Unlike traditional systems that allocate a fixed amount of memory regardless of the actual need, vLLM uses a dynamic memory allocation strategy. It allocates memory only when it’s needed and adjusts the allocation based on the system’s current workload. This prevents memory from sitting idle and ensures that no memory is wasted on tasks that don’t require it.

Cache Sharing Across Tasks:vLLM introduces the ability to share the KV cache across multiple tasks or requests. Instead of creating separate caches for each task, which can be memory-intensive, vLLM allows the same cache to be reused by different tasks. This reduces the overall memory footprint while still ensuring that tasks can run in parallel without performance degradation.

Handling Larger Batches:With efficient memory allocation and cache sharing, vLLM allows systems to process much larger batches of data at once. This is particularly useful in scenarios where processing speed and the ability to handle many requests at the same time are crucial, such as in large-scale AI systems or services that handle millions of user queries simultaneously.

Minimal Memory Waste:The combination of dynamic allocation and cache sharing means that vLLM can handle more tasks with less memory. It optimizes every bit of available memory, ensuring that almost none of it goes to waste. This results in near-zero memory wastage, which significantly improves system efficiency and performance.

Managing Limited Memory

When working with deep learning models, especially those that require significant memory for operations, you may encounter situations where GPU memory becomes insufficient. Two common techniques can be employed to address this issue: swapping and recomputation. Both methods allow for memory optimization, though they come with latency and computation time trade-offs.

1. Swapping

Swapping refers to the process of offloading less frequently used data from GPU memory to CPU memory when GPU resources are fully occupied. A common use case for swapping in neural networks is the KV cache (key-value cache), which stores intermediate results during computations.

When the GPU memory is exhausted, the system can transfer KV cache data from the GPU to the CPU, freeing up space for more immediate GPU tasks. However, this process comes at the cost of increased latency. Since the CPU memory is slower compared to GPU memory, accessing the swapped-out data requires additional time, leading to a performance bottleneck, especially when the data needs to be frequently swapped back and forth.

Advantages:

Saves GPU memory by offloading less essential data.

Prevents out-of-memory errors, allowing larger models or batch sizes.

Drawbacks:

2. Recomputation

Recomputation is another technique that helps conserve memory by reusing previously discarded data. Instead of storing intermediate activations (results from earlier layers of the model) during forward propagation, recomputation discards these activations and recomputes them on-demand during backpropagation. This reduces memory consumption but increases the overall computation time.

For instance, during the training process, the model might discard activations from earlier layers after they are used in forward propagation. When backpropagation starts, the model recalculates the discarded activations as needed to update the weights, which saves memory but requires additional computation.

Advantages:

Drawbacks:

Increases computation time since activations are recalculated.

May slow down the training process, especially for large and deep networks.

Conclusion

Determining the GPU memory requirements for serving LLMs can be challenging due to various factors such as model size, sequence length, and concurrent users. However, by understanding the different components of memory consumption—model parameters, KV cache, activations, and overheads—you can accurately estimate your needs.

Techniques like PagedAttention and vLLM are game-changers in optimizing GPU memory, while strategies like swapping and recomputation can help when facing limited memory.

FAQs

What is KV Cache in LLM inference?

The KV cache stores intermediate key-value pairs needed for generating tokens during sequence generation, helping models maintain context.

How does PagedAttention optimize GPU memory?

PagedAttention dynamically allocates memory in smaller, non-contiguous blocks, reducing fragmentation and improving memory utilization.

How much GPU memory do I need for a GPT-3 model?

GPT-3, with 175 billion parameters, requires around 350 GB of memory for weights alone, making it necessary to distribute the model across multiple GPUs.

What are the benefits of using vLLM?

vLLM reduces memory waste by dynamically managing GPU memory and enabling cache sharing between requests, increasing throughput and scalability.

How can I manage memory if I don’t have enough GPU capacity?

You can use swapping to offload data to CPU memory or recomputation to reduce stored activations, though both techniques increase latency.

Source link

Nextrope on Economic Forum 2024: Insights from the Event – Nextrope – Your Trusted Partner for Blockchain Development and Advisory Services

Web3

Metaverse Fashion Week Online

September 14, 2024

Nextrope on Economic Forum 2024: Insights from the Event – Nextrope – Your Trusted Partner for Blockchain Development and Advisory Services

Behavioral economics is a field that explores the effects of psychological factors on economic decision-making. This branch of study is especially pertinent while designing a token since user perception can significantly impact a token’s adoption.

We will delve into how token design choices, such as staking yields, token inflation, and lock-up periods, influence consumer behavior. Research studies reveal that the most significant factor for a token’s attractiveness isn’t its functionality, but its past price performance. This underscores the impact of speculative factors. Tokens that have shown previous price increases are preferred over those with more beneficial economic features.

Understanding Behavioral Tokenomics

Understanding User Motivations

The design of a cryptocurrency token can significantly influence user behavior by leveraging common cognitive biases and decision-making processes. For instance, the concept of “scarcity” can create a perceived value increase, prompting users to buy or hold a token in anticipation of future gains. Similarly, “loss aversion,” a foundational principle of behavioral economics, suggests that the pain of losing is psychologically more impactful than the pleasure of an equivalent gain. In token design, mechanisms that minimize perceived losses (e.g. anti-dumping measures) can encourage long-term holding.

Incentives and Rewards

Behavioral economics also provides insight into how incentives can be structured to maximize user participation. Cryptocurrencies often use tokens as a form of reward for various behaviors, including mining, staking, or participating in governance through voting. The way these rewards are framed and distributed can greatly affect their effectiveness. For example, offering tokens as rewards for achieving certain milestones can tap into the ‘endowment effect,’ where people ascribe more value to things simply because they own them.

Social proof, where individuals copy the behavior of others, plays a crucial role in the adoption of tokens. Tokens that are seen being used and promoted by influential figures within the community can quickly gain traction, as new users emulate successful investors. The network effect further amplifies this, where the value of a token increases as more people start using it. This can be seen in the rapid growth of tokens like Ethereum, where the broad adoption of its smart contract functionality created a snowball effect, attracting even more developers and users.

Token Utility and Behavioral Levers

The utility of a token—what it can be used for—is also crucial. Tokens designed to offer real-world applications beyond mere financial speculation can provide more stable value retention. Integrating behavioral economics into utility design involves creating tokens that not only serve practical purposes but also resonate on an emotional level with users, encouraging engagement and investment. For example, tokens that offer governance rights might appeal to users’ desire for control and influence within a platform, encouraging them to hold rather than sell.

Understanding Behavioral Tokenomics

Intersection of Behavioral Economics and Tokenomics

Behavioral economics examines how psychological influences, various biases, and the way in which information is framed affect individual decisions. In tokenomics, these factors can significantly impact the success or failure of a cryptocurrency by influencing user behavior towards investment

Influence of Psychological Factors on Token Attraction

A recent study observed that the attractiveness of a token often hinges more on its historical price performance than on intrinsic benefits like yield returns or innovative economic models. This emphasizes the fact that the cryptocurrency sector is still young, and therefore subject to speculative behaviors.

The Effect of Presentation and Context

Another interesting finding from the study is the impact of how tokens are presented. In scenarios where tokens are evaluated separately, the influence of their economic attributes on consumer decisions is minimal. However, when tokens are assessed side by side, these attributes become significantly more persuasive. This highlights the importance of context in economic decision-making—a core principle of behavioral economics. It’s easy to translate this into real-life example – just think about the concept of staking yields. When told that the yield on e.g. Cardano is 5% you might not think much of it. But, if you were simultaneously told that Anchor’s yield is 19%, then that 5% seems like a tragic deal.

Implications for Token Designers

The application of behavioral economics to the design of cryptocurrency tokens involves leveraging human psychology to encourage desired behaviors. Here are several core principles of behavioral economics and how they can be effectively utilized in token design:

Leveraging Price Performance

Studies show clearly: “price going up” tends to attract users more than most other token attributes. This finding implies that token designers need to focus on strategies that can showcase their economic effects in the form of price increases. This means that e.g. it would be more beneficial to conduct a buy-back program than to conduct an airdrop.

Scarcity and Perceived Value

Scarcity triggers a sense of urgency and increases perceived value. Cryptocurrency tokens can be designed to have a limited supply, mimicking the scarcity of resources like gold. This not only boosts the perceived rarity and value of the tokens but also drives demand due to the “fear of missing out” (FOMO). By setting a cap on the total number of tokens, developers can create a natural scarcity that may encourage early adoption and long-term holding.

Initial Supply Considerations

The initial supply represents the number of tokens that are available in circulation immediately following the token’s launch. The chosen number can influence early market perceptions. For instance, a large initial supply might suggest a lower value per token, which could attract speculators. Data shows that tokens with low nominal value are highly volatile and generally underperform. Understanding how the initial supply can influence investor behavior is important for ensuring the token’s stability.

Managing Maximum Supply and Inflation

A finite maximum supply can safeguard the token against inflation, potentially enhancing its value by ensuring scarcity. On the other hand, the inflation rate, which defines the pace at which new tokens are introduced, influences the token’s value and user trust.

Investors in cryptocurrency markets show a notable aversion to deflationary tokenomics. Participants are less likely to invest in tokens with a deflationary framework, viewing them as riskier and potentially less profitable. Research suggests that while moderate inflation can be perceived neutrally or even positively, high inflation does not enhance attractiveness, and deflation is distinctly unfavorable.

Source: Behavioral Tokenomics: Consumer Perceptions of Cryptocurrency Token Design

These findings suggest that token designers should avoid high deflation rates, which could deter investment and user engagement. Instead, a balanced approach to inflation, avoiding extremes, appears to be preferred among cryptocurrency investors.

Loss Aversion

People tend to prefer avoiding losses to acquiring equivalent gains; this is known as loss aversion. In token design, this can be leveraged by introducing mechanisms that protect against losses, such as staking rewards that offer consistent returns or features that minimize price volatility. Additionally, creating tokens that users can “earn” through participation or contribution to the network can tap into this principle by making users feel they are safeguarding an investment or adding protective layers to their holdings.

Social proof is a powerful motivator in user adoption and engagement. When potential users see others adopting a token, especially influential figures or peers, they are more likely to perceive it as valuable and trustworthy. Integrating social proof into token marketing strategies, such as showcasing high-profile endorsements or community support, can significantly enhance user acquisition and retention.

Mental Accounting

Mental accounting involves how people categorize and treat money differently depending on its source or intended use. Tokens can be designed to encourage specific spending behaviors by being categorized for certain types of transactions—like tokens that are specifically for governance, others for staking, and others still for transaction fees. By distinguishing tokens in this way, users can more easily rationalize holding or spending them based on their designated purposes.

Endowment Effect

The endowment effect occurs when people value something more highly simply because they own it. For tokenomics, creating opportunities for users to feel ownership can increase attachment and perceived value. This can be done through mechanisms that reward users with tokens for participation or contribution, thus making them more reluctant to part with their holdings because they value them more highly.

Conclusion

By considering how behavioral factors influence market perception, token engineers can create much more effective ecosystems. Ensuring high demand for the token, means ensuring proper funding for the project in general.

If you’re looking to create a robust tokenomics model and go through institutional-grade testing please reach out to contact@nextrope.com. Our team is ready to help you with the token engineering process and ensure your project’s resilience in the long term.

FAQ

How does the initial supply of a token influence its market perception?

The initial supply sets the perceived value of a token; a larger supply might suggest a lower per-token value.

Why is the maximum supply important in token design?

A finite maximum supply signals scarcity, helping protect against inflation and enhance long-term value.

How do investors perceive inflation and deflation in cryptocurrencies?

Investors generally dislike deflationary tokens and view them as risky. Moderate inflation is seen neutrally or positively, while high inflation is not favored.

Source link

How to Use Virtual Stores to Drive Holiday Shopping

Metaverse

Metaverse Fashion Week Online

September 13, 2024

How to Use Virtual Stores to Drive Holiday Shopping

It’s not surprising that holiday sales account for a significant amount of retailers’ annual sales. Brands are continually searching for engaging and entertaining campaigns to capitalize on drive holiday shopping and sales. A 3D virtual store offers a new experiential shopping format and serves as a viable sales channel. With a virtual experience, brands can create an environment that is modeled off of their retail store, a creative concept store or a completely fantastical location.

Virtual holiday stores allow brands to create completely immersive and discovery-driven experiences packed with educational content, primed for the busy gifting season. With 3D content and innovative features such as gamification and gift recommendation quizzes, brands across verticals were able to increase engagement, conversion and reach during the busiest time of year.

Key benefits of virtual holiday stores include:

A new and engaging virtual channel that stands out from other holiday campaigns

Interactive and discovery-driven shopping experience that drives sales and marketing ROI

Engaging features that connect consumers to the brand

A one-stop shop to discover, research and purchase gifts with seamless e-commerce integration

Lower production costs for CGI visuals compared to on-location shoots and brick-and-mortar displays

To learn more about how immersive experiences impact holiday sales, download our e-book comparing industry benchmarks from partner Salesforce to Obsess first-party data.

Examples of Virtual Holiday Stores

Obsess has worked with a number of brands and retailers to create memorable holiday virtual stores and experiences. Each experience featured fantastical displays and larger-than-life decorations, while maintaining core e-commerce functionality for an immersive and enriching online shopping experience that drives virtual holiday shopping. From brands in fashion and beauty to CPG and beyond, these unique e-commerce experiences are able to immerse holiday shoppers into a virtual world unique to the brand.

Here are 5 examples of Obsess-powered virtual holiday experiences:

J.Crew Virtual Ski Chalet

J Crew holiday Virtual store shopping powered by Obsess

J.Crew launched a digital après-ski retreat to delve into the brand’s essence and seasonal offerings. Full of interactive elements and holiday-inspired ambiance, the innovative ski chalet merchandised the latest men’s and women’s collection in unique ways. Users would navigate to the Snow Lodge for the newest women’s styles, and the Ski Chalet for men’s—a curated selection of seasonal looks and holiday gifts were seamlessly woven into the experience, offering a fully shoppable journey that encouraged virtual holiday shopping.

To add an extra layer of excitement throughout the experience, the brand included dynamic animations and custom gamification that engaged shoppers and deepened consumer connections. Festive gamification included a scavenger hunt game, vintage video game-style Ski Racing Game and a branded quiz that tested shopper’s knowledge on the brand’s long and storied heritage. The scavenger hunt also unlocked a chance at an exclusive luxe cashmere giveaway, which prompted users to explore the entire store. Click to read more about the Holiday World of J.Crew.

Corona Holiday Beach Hut

Constellation Brand’s Corona invited new and longtime fans to Feliz Navi-Drip it like it’s hot with Snoop Dogg in their Holiday Beach Hut experience. The Mexican beer brand revitalized one of their famed holiday commercials into an immersive experience that transported visitors to the same palm tree-filled beach sunset environment. The experience amplified their holiday campaign in VR and enabled consumers to explore products and engage with the brand in new ways.

Fans dived into the holiday spirit with interactive features throughout the experience, such as a custom arcade-style ornament collection game. In the colorful string light laden beach hut, visitors were able to watch the aforementioned commercial and hang with brand ambassador Snoop Dogg. Using AR technology, visitors were able to take a picture with Snoop and create a custom, shareable holiday card. Visitors were also given a chance to match Snoop’s iconic look if they entered the sweepstakes to win a limited-edition Talia Coles robe. Read more about the virtual experiences Obsess has created with Corona, including the Holiday Beach Hut.

Sephora X Cosmopolitan Gifting Suite

Sephora x Cosmopolitan Virtual Gifting Experience Powered by Obsess

Renowned Hearst publication Cosmopolitan collaborated with beauty retailer, Sephora, and released the Gifting Suite. The virtual holiday experience empowered users seeking a scent for personal indulgence or as a sophisticated gift to discover unique scent profiles for their holiday shopping needs. Redefining the fragrance content and online shopping experience, the immersive experience focused around personalized luxury and discovery.

Lined with chic twinkly lights and festive disco ball-covered trees, the suite featured a curated selection of scents directly shoppable in the experience and thoughtfully arranged by notes. Central to the experience was a Cosmo-esque quiz, where shoppers were invited to unravel the nuances of their fragrance preferences. Click to read more about the Sephora X Cosmopolitan Gifting Suite.

H-E-B Grocery Holiday Home

Step into the Holiday Home of beloved Texas-based grocery chain, H-E-B. An immersive wonderland of flavors, recipes and holiday products unfolded through interactive features and cheerful animations. Adorned with festive decorations, the Texas-style home featured multiple rooms ready for the holidays. Users were greeted by virtual chef avatars that guided shoppers through classic recipes, covering appetizers, mains, sides and desserts. Each recipe pop-up was complemented by shoppable links of H-E-B products, offering a seamless buying journey that allowed users to kick-start their holiday cooking.

The virtual experience also provided the brand with a platform to spotlight its philanthropic endeavors and rich history. An interactive scavenger hunt game engaged users through a discovery-driven educational adventure that unveiled facts about the brand and unlocked exclusive holiday deals. Read more about the virtual grocery shopping experience from H-E-B Grocery.

Laneige Virtual World

Korean beauty brand, Laneige, launched a dreamy virtual world that enabled shoppers to explore fan-favorite products and their brand story in a fully immersive context. The experience launched with discovery-driven games and a variety of rooms dedicated to different product lines, and enabled shoppers to interact with products in different ways, including 3D views and multimedia content around ingredients and application. The brand regularly updated the cloudscape to align with their marketing calendar, including a full seasonal refresh that reflected new offerings and product releases for the 2023 holiday season. Festive updates included limited-edition collections, new flavors of the iconic Lip Sleeping Mask and a revamp of the classic scavenger hunt game.

“The virtual store is a powerful customer engagement tool that will allow shoppers to dive into Laneige’s unique scientific expertise and entertaining digital content, while also helping them select the right product for their skin needs, including our bestselling Water Sleeping Mask and Lip Sleeping Mask,” quoted Julien Bouzitat, Chief Marketing and Digital Officer of Laneige parent company Amorepacific U.S. Read more about the Laneige Virtual World, and it’s continued evolution.

These examples only scratch the surface of the limitless possibilities that can go into a digital holiday experience. To find out more on how you can create your own holiday virtual experience, book a demo or email us at contact@obsessvr.com.

Source link

OpenAI Launches GPT-01 ‘Strawberry’: A Breakthrough in Advanced Reason

Web3

Metaverse Fashion Week Online

September 13, 2024

OpenAI Launches GPT-01 ‘Strawberry’: A Breakthrough in Advanced Reason

On September 12, 2024, OpenAI introduced GPT-01, also known as “Strawberry.” This release is the first of a planned series of reasoning models aimed at solving complex tasks more effectively. While previous models, like GPT-4, excelled at generating human-like text and responses, GPT-01 brings something new: advanced reasoning capabilities to tackle multi-step problems, such as complicated math, logic, and coding tasks.

What Is Advanced Reasoning in AI?

Advanced reasoning refers to the AI’s ability to process information logically and step-by-steply, similar to how a human would tackle a problem. While earlier models mainly focused on pattern recognition and data-based predictions, GPT-01 can break down tasks into smaller steps, analyze them, and provide a coherent solution.

This leap in reasoning power makes GPT-01 particularly effective in areas where multi-step logic is essential, such as coding, mathematical proofs, and strategic planning.

GPT-01 represents a significant milestone in AI research, particularly in reasoning. By pushing the boundaries of what AI can achieve, this model sets the stage for future innovations in AI development, bringing us closer to creating autonomous systems capable of complex decision-making.

The introduction of advanced reasoning in GPT-01 paves the way for more sophisticated AI applications, particularly in fields that require logical thinking and problem-solving. As models evolve, AI may take on more significant roles in decision-making processes, from healthcare to engineering.

Why Is the Model Called “Strawberry”?

GPT -01 was nicknamed “Strawberry” to reflect its user-friendliness and adaptability. According to OpenAI, the model was designed with human-like interaction in mind, making it a more intuitive and collaborative tool.

Key Features of GPT-01 “Strawberry”

1. Coding: Thoroughly Analyzing Each Instruction

It’s fascinating how GPT-01 generates code compared to GPT-4. The GPT-01 preview version takes time, carefully considering the prompt. In programming, we often provide detailed instructions, and GPT-4 tends to miss or overlook some aspects, much like how we might feel when juggling too many tasks simultaneously. However, GPT-01 meticulously processes all the information, analyzing every requirement slowly and thoroughly.

In a demonstration, they used the following coding prompt:

2. Reasoning: Understanding Context and Surroundings

GPT-01 is designed to tackle common-sense reasoning, where most large language models (LLMs) struggle. It can make decisions in complex situations, such as identifying relationships between objects and their physical context.

In the demonstration, they used this prompt:

3. Mathematics: Tackling Complex Problems

In the demonstration, GPT-01 handled math problems of medium difficulty easily. It efficiently processed tasks involving logical sequences, groupings, and trends.

The math prompt presented in the video was:

Comparing GPT-01 to GPT-4

Although GPT-01 is slower and more expensive to use than GPT-4, it excels in complex reasoning tasks. In a test against the International Mathematics Olympiad’s qualifying exam, GPT-01 correctly solved 83% of the problems, compared to GPT-4’s 13%.

Here’s a comparison chart highlighting the key differences between GPT-01 (Strawberry) and GPT-4:

FeatureGPT-01 (Strawberry)GPT-4

Release DateSeptember 2024March 2023

Core FocusAdvanced reasoning and multi-step problem-solvingGeneral-purpose language generation

Reasoning CapabilitiesSuperior in complex tasks like coding, math, and logical reasoningModerate reasoning skills

SpeedSlower, takes more time to process multi-step tasksFaster response time for general queries

Cost (per 1M input tokens)$15$5

Cost (per 1M output tokens)$60$15

AccuracyHigher accuracy in reasoning tasks, fewer hallucinationsAccurate for general text generation but prone to more hallucinations in complex tasks

Use CasesBest suited for math, coding, logic, and strategic tasksIdeal for text generation, creative writing, and casual Q&A

Complex Problem SolvingExcels in multi-step reasoning, performs well on math exams and programming tasksLimited, struggles with advanced multi-step problems

Conversational Context RetentionRetains context over extended dialogues effectivelyAdequate but can lose context in long conversations

Factual KnowledgeLess proficient at factual world knowledgeStronger at handling factual information

Browsing and File ProcessingDoes not support browsing or file/image processingCan browse the web (with plugins) and process files (with plugins)

Target AudienceDevelopers, engineers, educators, researchersGeneral users, content creators, casual inquiries

Training ApproachTrained with reinforcement learning for reasoning tasksTrained on large datasets for language prediction

AvailabilityCurrently available to ChatGPT Plus and Team usersWidely available for both free and paid users

Use Cases for GPT-01

Coding and Programming

GPT-01 significantly outperforms its predecessors when it comes to programming tasks. It can process complicated code, understand step-by-step instructions, and produce real-time error-free outputs, making it ideal for developers and engineers.

Mathematical Problem Solving

The model’s ability to solve complex math problems is a notable advancement. For example, GPT-01 can tackle multistep word problems and logical puzzles, making it a valuable tool for anyone studying or working in math-intensive fields.

Business Applications

In business, GPT-01 can assist with data analysis, risk assessment, and long-term strategic planning by logically processing incomplete data and offering suggestions based on trends and predictions.

Education and Tutoring

With its ability to break down complex problems and offer step-by-step reasoning, GPT-01 can act as a student tutor. Whether in math, coding, or philosophy, this model can offer detailed explanations and help learners understand difficult concepts.

Challenges with GPT-01

Despite its advancements, GPT-01 does have limitations. The model is slower than GPT-4 and requires more computational power, which increases its cost. It also struggles with factual knowledge and cannot browse the web or process files and images, limiting its use in certain scenarios.

What’s Next for OpenAI?

OpenAI has indicated that GPT-01 is just the beginning of a new series of reasoning models. The company also focuses on improving the model’s speed, cost-efficiency, and factual accuracy in future iterations. As these models evolve, they will likely take on more advanced tasks, moving closer to human-like intelligence.

Conclusion

OpenAI’s GPT-01 “Strawberry” is a groundbreaking development in AI, marking the first model with advanced reasoning capabilities. While slower and more expensive than previous models, GPT-01 excels at solving complex problems like coding, math, and logical reasoning. As AI continues to evolve, GPT-01 paves the way for smarter, more collaborative AI systems that can assist humans in more nuanced and sophisticated ways.

FAQs

What makes GPT-01 different from GPT-4? GPT-01 is designed for complex reasoning tasks, such as coding and mathematical problem-solving, while GPT-4 is faster and better suited for general text generation.

Why is GPT-01 called “Strawberry”? The name reflects its user-friendly nature and adaptability, emphasizing its collaborative potential.

How much does GPT-01 cost to use? GPT-01-preview costs $15 per 1 million input tokens and $60 per 1 million output tokens, making it more expensive than GPT-4.

Can GPT-01 browse the web? No, GPT-01 does not currently have browsing capabilities or the ability to process files and images.

What industries will benefit most from GPT-01? Industries like coding, education, business analysis, and healthcare will see the most immediate benefits from GPT-01’s reasoning abilities.

Source link

Natural Makeup Looks: How to Achieve a Flawless ‘No-Makeup’ Makeup – Hautelist

Couture

Metaverse Fashion Week Online

September 13, 2024

Natural Makeup Looks: How to Achieve a Flawless ‘No-Makeup’ Makeup – Hautelist

The allure of the ‘no-makeup’ makeup look lies in its simplicity and elegance. It’s about enhancing your natural beauty, creating a fresh, radiant appearance that seems effortless. However, achieving this look requires more than just minimal makeup; it’s about selecting the right products and applying them skillfully. Here’s how to master the art of natural makeup for a flawless, everyday look.

1. Start with Skincare

The foundation of any great makeup look is healthy skin. A natural makeup look requires a well-moisturized, smooth canvas. Begin by cleansing your face to remove any impurities. Follow up with a hydrating moisturizer suited to your skin type to ensure your skin is supple and hydrated. If your skin tends to be dry, opt for a richer cream, while those with oily skin may prefer a lightweight, oil-free moisturizer.

Shop now

Don’t forget to apply sunscreen, even if you’ll be indoors. Sunscreen protects your skin from harmful UV rays and prevents premature aging, which is crucial for maintaining a fresh and youthful complexion.

Shop now

2. Prime for Perfection

Shop now

Primers can be your best friend when aiming for a natural look. A good primer smooths out your skin, blurs imperfections, and helps your makeup last longer. Choose a primer that suits your skin type; mattifying primers are great for oily skin, while illuminating primers add a subtle glow to dry or dull skin.

3. Keep It Light with Foundation

Shop now

For a natural look, less is more when it comes to foundation. Opt for a lightweight, sheer foundation or a tinted moisturizer that evens out your skin tone without masking your natural features. BB creams or CC creams are also excellent choices as they provide coverage while maintaining a natural finish.

Apply the foundation sparingly, focusing on areas where you need coverage, such as the T-zone, and blend well with a damp beauty sponge or brush for a seamless finish. If you have clear skin, you might skip foundation altogether and just use a concealer where needed.

4. Conceal Strategically

Concealer is essential for a natural makeup look, but the key is to use it strategically. Choose a concealer that matches your skin tone exactly or is just a shade lighter. Apply it only on areas that need extra coverage, such as under the eyes, around the nose, or on any blemishes.

Shop now

For a bright-eyed look, apply a small amount of concealer in the inner corners of your eyes and blend it outwards. This will instantly lift and brighten your eyes without looking overdone.

5. Subtle Definition with Brows

Well-groomed brows frame your face and are crucial for a polished look. Use a brow pencil or powder that matches your natural brow color to fill in any sparse areas with light, feathery strokes. The goal is to enhance the shape of your brows while keeping them looking soft and natural. Finish with a clear or tinted brow gel to set them in place.

6. Enhance Your Eyes Softly

For the eyes, stick to neutral shades that enhance your natural eye color. A soft taupe or beige eyeshadow on the lids, blended well, adds depth without being obvious. You can define your lash line by applying a thin line of brown or black eyeliner close to your lashes, then smudge it slightly for a softer look.

Mascara is a must, but keep it minimal. One or two coats of a lengthening mascara will open up your eyes without making your lashes look clumpy. If you prefer a more natural look, opt for a brown mascara instead of black.

7. Add a Natural Flush

For a natural flush, cream blushes work best as they blend seamlessly into the skin, giving you a healthy, dewy glow. Choose a blush shade that mimics your natural flush—think soft pinks, peaches, or corals—and apply it to the apples of your cheeks. Blend well upwards towards your temples for a lifted effect.

8. Luminous Lips

Finish your look with a lip color that enhances your natural lip shade. Tinted lip balms, sheer lipsticks, or nude lip glosses are perfect for this. They add a hint of color and moisture without overpowering your look. If you prefer a matte finish, opt for a nude lip liner all over the lips, topped with a balm to keep them hydrated.

Shop now

9. Set with a Mist

Shop now

To ensure your makeup stays put all day while retaining its natural finish, set it with a hydrating facial mist or a lightweight setting spray. This step will also help melt the makeup into your skin, giving it a fresh, dewy finish.

The ‘no-makeup’ makeup look is about enhancing your natural beauty with subtle, well-applied products. By focusing on skincare, using light, natural-toned makeup, and applying it strategically, you can achieve a flawless, radiant look that appears effortlessly chic. Whether for daily wear or a special occasion, this look is timeless and versatile, perfect for anyone who wants to feel confident and beautiful in their own skin.

Source link

Can Machine Learning Models Learn from Each Other?

Web3

Metaverse Fashion Week Online

September 12, 2024

Can Machine Learning Models Learn from Each Other?

Artificial Intelligence (AI) is evolving rapidly, particularly in how machine learning (ML) models learn and interact with one another. Traditional methods heavily depend on vast human-generated datasets, but recent advancements explore the idea of models teaching each other, resembling human learning dynamics. This approach, which uses synthetic data and novel learning paradigms, has significant implications for overcoming data scarcity, enhancing accessibility, and democratizing AI.

The Rise of Collaborative Learning in AI

Artificial intelligence has reached new heights with the development of collaborative learning techniques, where machine learning (ML) models learn from one another, mimicking human-like interactions. This approach moves away from traditional reliance on vast amounts of human-generated data, which is becoming harder to obtain and more expensive. Instead, AI models now generate synthetic data and use it to improve themselves and others in an efficient, iterative process. This shift could democratize AI, making advanced technology more accessible to small companies and individuals with limited resources.

The Role of Synthetic Data in Transfer Learning

At the core of collaborative learning is synthetic data generation, which allows models to create their own datasets rather than relying solely on human-generated data. This synthetic data is a crucial component of transfer learning, where a larger, more capable model acts as a teacher to a smaller, less powerful model. This teacher-student dynamic enables the smaller model to gain insights from the larger one without extensive retraining on expensive and scarce datasets.

Transfer learning with synthetic data transforms how models are trained, making it possible for even small-scale AI projects to benefit from the knowledge embedded in larger models. Two key projects, Self-Instruct and Alpaca, have demonstrated the immense potential of this approach.

Self-Instruct: Enhancing Models with Self-Generated Tasks

Self-Instruct is an innovative project that enhances a model’s instruction-following capabilities by allowing it to train on synthetic tasks it creates. The process, called iterative refinement, allows the model to generate instructions, perform tasks, and then evaluate and refine its own outputs.

In the Self-Instruct project, a pre-trained language model generates tasks that consist of instructions, contextual inputs, and expected outputs. These tasks are not manually crafted by humans but produced by the model itself. As the model continues to work through these tasks, it improves its ability to follow instructions in new, unseen scenarios. Notably, a quality filter ensures that only the best-generated tasks are used for further training, eliminating low-quality or irrelevant data.

The iterative nature of this refinement allows the model to continuously improve, making it increasingly adept at understanding and executing instructions with minimal human intervention. By reducing dependency on human-generated data, Self-Instruct offers a way to create highly specialized models that can perform well across various tasks without requiring vast datasets or retraining.

Alpaca: Efficient AI with Teacher-Student Dynamics

While Self-Instruct focuses on self-generated learning, Alpaca introduces a hierarchical approach, where a more advanced model teaches a smaller, less capable one. Alpaca leverages OpenAI’s text-davinci-003, a highly advanced language model, to create instruction-following tasks for LLaMA 7B, a smaller and less resource-intensive model developed by Meta.

The Alpaca project demonstrates that it is possible to train smaller models to follow complex instructions effectively, even with limited resources. This teacher-student setup replicates a real-world classroom scenario, where the more knowledgeable teacher provides guidance and tasks that help the student model learning. The student model, in turn, benefits from the teacher’s advanced capabilities, achieving a high level of performance without requiring the same level of computational power or data as the teacher.

This method reduces the cost of training AI models and allows smaller entities to participate in AI development. Companies and individual researchers with limited budgets can still produce AI models that perform well on specific tasks, such as natural language understanding or instruction-following.

How Collaborative Learning is Democratizing AI

The success of projects like Self-Instruct and Alpaca highlights the potential for collaborative learning to democratize access to advanced AI technology. Previously, training large models required immense computational resources and datasets, making it the domain of large tech companies with deep pockets. However, with collaborative learning techniques, smaller models can tap into the knowledge of larger ones, drastically reducing the need for resources.

Critical Benefits of Collaborative Learning:

Resource Efficiency: Smaller models can achieve high performance by learning from larger models, reducing the need for extensive computational power and large datasets.

Cost Reduction: Companies no longer need to invest heavily in hardware or data acquisition. Models like Alpaca prove that effective instruction-following models can be built on a budget.

Scalability: The teacher-student framework can be applied across industries, allowing for scalable AI development even in resource-constrained environments.

Increased Accessibility: Collaborative learning lowers the barriers to AI development, ensuring that smaller companies and individual researchers can contribute to the AI landscape.

The Future of Collaborative AI

As AI continues to evolve, the future will likely see an increasing reliance on collaborative learning methods. These techniques solve the pressing issue of data scarcity and introduce a more sustainable way to train and refine models. By utilizing teacher-student setups and synthetic data generation, the AI field is set to become more inclusive, with even small players able to develop powerful models for various applications.

The combination of transfer learning, iterative refinement, and hierarchical teaching structures presents an exciting future for AI development. Whether in healthcare, finance, customer service, or other industries, collaborative learning will enable more entities to harness the power of AI, regardless of their size or resources.

In essence, collaborative learning redefines how AI models are trained and deployed, making them more efficient, cost-effective, and accessible. This paradigm shift not only accelerates the pace of innovation in AI but also ensures that the benefits of these advancements are widely shared, empowering businesses and individuals alike to explore new frontiers in machine learning.

Learning Paradigms: Beyond Traditional Fine-Tuning

Machine learning has come a long way since the early days of simple algorithms and data processing. Today, researchers are exploring advanced learning paradigms that push the boundaries of what AI can achieve. These new methods go beyond traditional fine-tuning, offering innovative ways for models to learn, adapt, and improve without the typical limitations of human-generated data and extensive retraining. Let’s dive into three key approaches—iterative Refinement, Knowledge Distillation, and Learning by Teaching—that are reshaping the landscape of collaborative machine learning.

Iterative Refinement is a process where models enhance their capabilities by generating and learning from their outputs. Unlike conventional training, which relies heavily on external datasets, Iterative Refinement creates a feedback loop where the model continuously assesses and improves itself.

Here’s how it works: the model generates synthetic data—often tasks or scenarios similar to what it’s already been trained on—and then uses these outputs as new training data. By iterating on this process, the model identifies errors, adjusts its parameters, and refines its performance over time. This approach is similar to a student reworking practice problems to master a concept, constantly learning from mistakes and refining their understanding.

Benefits:

Data Efficiency: Models can continue to learn without needing vast amounts of new, human-generated data.

Self-Sufficiency: The model independently identifies areas of improvement, leading to continuous growth.

Customization: Models can adapt to specific tasks, making them highly specialized without needing extensive retraining.

Challenges:

Quality Control: It is crucial to ensure that the synthetic data generated is of high quality; poor-quality data can lead to suboptimal learning.

Computational Resources: Iterative processes can be computationally expensive, requiring significant processing power and time.

2. Knowledge Distillation: Transferring Wisdom from Teacher to Student

Knowledge Distillation is inspired by the traditional educational model, where a knowledgeable teacher guides students through complex subjects. In this machine-learning context, a large, powerful model (the teacher) transfers its expertise to a smaller, less capable model (the student). This transfer allows the student model to perform highly without extensive retraining on large datasets.

The process involves the teacher model providing output predictions, which the student model then learns to replicate. Through this method, the student model not only learns the task but also picks up the teacher’s nuanced decision-making processes, resulting in a highly efficient and specialized model that performs well with fewer resources.

Benefits:

Resource Efficiency: The student model is much smaller and requires fewer computational resources, making it suitable for deployment in resource-constrained environments.

Performance Enhancement: The student model can perform at levels comparable to the larger teacher model, making high-quality AI accessible even without powerful hardware.

Scalability: This approach allows AI capabilities to be scaled down effectively, enabling practical applications in various industries.

Challenges:

Loss of Detail: The student model may lose some of the finer points of the teacher’s knowledge, potentially leading to reduced performance in highly complex tasks.

Dependency on Teaching Quality: This approach’s effectiveness depends heavily on the quality of the teacher model and the training process.

3. Learning by Teaching: A Two-Way Feedback Loop for Continuous Improvement

Learning by Teaching draws inspiration from a well-known human learning strategy: teaching others. This approach introduces a unique feedback loop where a student model doesn’t just learn passively but actively contributes to the teacher model’s improvement. Here’s how it works: after the teacher model instructs the student, the student’s performance is assessed, and this feedback is returned to the teacher. The teacher model then uses this performance data to adjust its teaching strategy, creating a dynamic and reciprocal learning environment.

This method mirrors how a tutor learns to refine their explanations based on how well their student understands and applies the material. In machine learning, this approach allows models to enhance their teaching methods, adapting and evolving to improve.

Benefits:

Enhanced Learning Efficiency: The feedback loop ensures that the teacher and student models continuously improve, leading to more effective learning outcomes.

Adaptive Teaching: The teacher model can adjust its approach based on real-time student performance data, leading to a more personalized learning experience.

Innovation in Learning Paradigms: This method introduces interactivity and adaptability beyond traditional model training techniques.

Challenges:

Complex Implementation: Setting up a feedback loop between teacher and student models requires sophisticated programming and robust evaluation metrics.

Risk of Overfitting: The teacher model might overly tailor its teaching to the student’s current performance, potentially limiting the broader applicability of the learned knowledge.

The potential applications of models learning from each other are vast and varied. From creating specialized models that perform specific tasks without needing large, general-purpose datasets to address privacy and copyright concerns, these techniques can redefine AI training.

However, challenges remain. For example, fine-tuning can lead to “catastrophic forgetting,” where a model loses its general abilities in favor of specialized skills. This trade-off raises questions about the broader applicability of collaborative learning methods beyond specific use cases.

Philosophical Implications: Machines Learning from Machines

The concept of machines teaching machines carries philosophical intrigue. It parallels human learning, where teaching often solidifies one’s understanding. Richard Feynman, a physicist known for his teaching techniques, highlighted the power of learning by teaching—a principle that now finds resonance in machine learning. The emerging AI techniques reflect a creative and experimental spirit, pushing the boundaries of how models can evolve autonomously.

Final Thoughts

Collaborative learning through synthetic data and model interaction is a promising avenue in AI research. It challenges traditional data-centric paradigms, offering new ways to train models efficiently and make AI accessible to a wider audience. While the full potential of these methods is still unfolding, they represent a significant step towards more flexible, resource-efficient, and innovative AI systems.

This exploration of machine models learning from each other underscores a pivotal shift in AI development. It embraces the creativity and ingenuity of human-inspired learning techniques, making the future of AI both exciting and accessible.

Source link

Christopher Ortiz AKA kiririn51 Talks .45 PARABELLUM BLOODHOUND, Inspirations, Fan Reactions, VA-11 Hall-A, The Silver Case, and Much More – TouchArcade

Gaming

Metaverse Fashion Week Online

September 12, 2024

Christopher Ortiz AKA kiririn51 Talks .45 PARABELLUM BLOODHOUND, Inspirations, Fan Reactions, VA-11 Hall-A, The Silver Case, and Much More – TouchArcade

Over the years, I’ve been able to interview some of my favorite developers ever including a few I never though would be possible, but it isn’t often I get to talk to one of the few people responsible for what is likely my favorite game of all time. That’s where Christopher Ortiz AKA kiririn51 of Sukeban Games comes into the picture. We’ve covered their titles on TouchArcade for years because at one point VA-11 Hall-A was even supposed to come to iPad (and I asked about that later in this interview). With Sukeban Games’ newest project, .45 PARABELLUM BLOODHOUND, officially announced, I had a chance to have a long chat with Christopher Ortiz about the game, fan reactions, VA-11 Hall-A, inspirations, Suda51, The Silver Case, and also coffee of course.

TouchArcade (TA): Tell us a little bit about yourself and what you do at Sukeban Games.

Christopher Ortiz (CO): I’m Chris, I’m a game creator and I do way too much in this company. I like to hang out with friends and eat delicious food when I’m not locked in.

TA: I last spoke to you in 2019 around the time VA-11 Hall-A hit PS4 and Switch following its prior release on PS Vita and its debut on PC. Even back then, as a fan of the game, it was wild seeing so much merchandise and promotion in Japan. You recently visited Japan for Bitsummit. How has it been for you visiting Japan and seeing the reception to VA-11 Hall-A and now your new project, .45 PARABELLUM BLOODHOUND?

CO: Japan is like my home away from home even if its government doesn’t like it, so that’s a homecoming for me. Very emotionally charged. I also haven’t been at a game event as an exhibitor since Tokyo Game Show 2017; that was 7 years ago. That’s 7 years of roaming game events and being like “I want that” “I want some of this energy”. Now I feel like a pro-wrestler that’s coming out of retirement to a brand new world and industry; I’m lost, confused, not knowing if I still have it… but I was worried about nothing. People didn’t forget about us and still support the studio no matter what, so I’m never taking shit for granted, or letting these memories slip away. This will be my fuel to move forward.

TA: I consider VA-11 Hall-A one of my favorite games of all time, and I replay it each holiday season. When you worked on it many years ago, did you ever expect to see it grow this big and even get multiple figures with a new one coming for Jill soon?

CO: I didn’t expect the game to sell more than 10-15k copies, but we definitely knew we had something special brewing, or else we wouldn’t have pushed through with it. It’s just that the scale of said success was a little overwhelming and I think we’re still recovering from some of its unexpected side-effects.

TA: VA-11 Hall-A is now playable on PC, Switch, PS Vita, PS4, PS5 (through back compatibility). Whatever happened to that iPad version that was announced years ago? Are ports like that up to Ysbryd or do you also have some involvement? I’d love to see it hit Xbox as well if that’s possible.

CO: I actually playtested a build for iPad but it didn’t go anywhere for whatever reason. Maybe I failed to respond to an email. You’d have to ask the publisher.

TA: Many years ago, Sukeban Games were just Kiririn51 (yourself) and IronincLark (Fer). How has the team changed since then?

CO: We’re currently six people. Some have come and go, but overall we want to keep a small and tight operation.

TA: Leading into this question -> How has it been working with MerengeDoll?

CO: Merenge’s a trooper. She has this supernatural ability to pull ideas directly out of my brain and give them visual shape, so it’s always been a pleasure to work with her. It sucks that some of the projects she was working on as main artist got canned for reasons beyond our power, but it is what it is. The day will come when people will get to see Merenge’s true power. There’s still a lot of Merengedoll in 45pb so that’s cool as hell.

TA: Can you talk about how it was working with Garoad on the music on VA-11 Hall-A? Just like the game, the soundtrack is one of my favorites ever.

CO: Michael and I always had similar music tastes and influences, so the process was very free-form. He’d make a track, and I’d love the shit out. Repeat the process until the soundtrack is ready to go. Sometimes I would send him a song I liked as a reference, sometimes he’d create a completely original song that would inspire images in the game, then the images would inspire more music. This synergy gave the game a firm identity that stands the test of time in my opinion.

TA: I didn’t realize it until recently, but VA-11 Hall-A kind of became one of those indie games that had a very vocal fanbase and got a good amount of merch that keeps selling out. I think the vinyl box sets are also on multiple pressings now and that SLUT shirt keeps selling out. A good friend of mine recently bought it as well. How much input do you have in the merchandise? Is there anything you want to see made that hasn’t happened yet?

CO: I don’t have much input on merch creation. I mostly give the thumbs up or down once a whole chain of people already made the hard choices beforehand. Would like to be a little more involved for 45pb now that I know what I’m dealing with.

TA: I still have a few more questions before getting to your new game. Let’s go back to 2019 for a second. Playism’s JP release of VA-11 Hall-A included a fantastic art book cover. I really wish I could get that piece of art signed by you and framed. Can you talk about the inspiration for it and how you pay homage to your favorites like that in your work?

CO: Back when I drew that cover I was going through really tough times; even though I hadn’t fully realized yet. We were very focused on surviving the collapse of our country and many other things.

During this time, at our old office, we’d listen to a lot of Gustavo Cerati; his album Bocanada to be more precise and its tunes managed to keep us going in the face of uncertainty. So when I was asked to draw a new piece for the artbook I couldn’t help but pay homage to it. I admit now that it’s a little too overt and I would do it differently now, but I’m still proud of it. In fact, I’ve been reconsidering my approach to inspirations a lot in the past few years, and this will be obvious once people play 45pb.

TA: You and Fer have spoken a lot about VA-11 Hall-A over the years, but I cannot pass up a chance to talk about how amazingly written and designed the characters are. Looking back at the work you put into it, did you expect certain characters to become as popular as they did?

CO: I expected Stella to be the most popular one since her gifs would often go viral before release, but you can never predict this sort of thing. And it’s like I said before; I sort of knew certain things were gonna be a hit, I’m just unable to properly articulate why I thought that way or why it worked, and at this point I rather not know. The moment a hunch becomes a science the magic disappears. Formulas are terrible in this line of work. You gotta let things flow; Let them become their own beautiful thing.

TA: I’ve joked with friends about how N1RV Ann-A is my “Silksong”, but I have no problem waiting as long as it takes. I still enjoy revisiting VA-11 Hall-A often. Do you go back and try out whatever you had done for N1RV Ann-A or VA-11 Hall-A while working on unrelated projects?

CO: I like jotting down lore and characterization stuff for a rainy day. I love to draw Sam, I love coming up with new designs, characters; playing around with the general looks of the game, ideas for shots during cinematics; one liners, environments, atmosphere. I even like to imagine “what if this didn’t have to be a bartending game?” and other extreme musings, but that’s about the extent of the mindspace I dedicate to it currently.

Once we’re done with 45pb Nirvana’s development will pick up dramatically, though that really depends on if the spark is still there by then. So far it has no signs of extinguishing.

TA: As a huge Suda fan myself, I’m curious what you thought of No More Heroes 3 and Travis Strikes Again? As much as I love No More Heroes 3 (more than 11 playthroughs), I think Travis Strikes Again might just be Suda’s most “Suda” game yet.

CO: I REALLY like No More Heroes 3’s combat, but I wasn’t a fan of its writing. Maybe it was COVID, and game development is really tough; especially with the hardcore deadlines they had. But in general I think it wanted to be one thing at the start and then had to become another in order to be shipped. It’s a shame, and it is what it is. I just hope they only make new original games from now on and forget about sequels and reboots. Re-releases are OK; especially of the lost media kind, like Frog Minutes or the 25th Ward back in the day.

As for Travis Strikes Again, I agree it’s the best of the newer games. Feels like reading someone’s diary sometimes and that’s the kinda art I enjoy. I guess I didn’t see much of that in 3 outside of continuing some plot threads from TSA.

TA: What are your thoughts on Grasshopper Manufacture under Netease and the remasters announced? Suda even mentioned today that he’d love to bring Flower Sun and Rain to Steam.

CO: Netease is a big corporation so my hope would be that Grasshopper gets all the money and time they need to cook.

TA: VA-11 Hall-A’s journey from PC to PS Vita was quite something, and it involved many parties across regions. I remember buying the Japanese release just for the box art with no English included for publisher reasons. When it came to Switch and PS4, I think you wanted the JP release to have English so anyone could import it. How has it been for you now in Argentina trying to get your own game’s merchandise and merchandise in general? It is a pain over here with delays and extra import fees.

CO: I simply don’t import anything these days. I don’t wanna bother with Argentinian customs. Protectionist policies are stupid. Sure, make all electronics more expensive to import so the local market has a chance to compete, but there’s no Argentinian PlayStation to compete with, is there? There’s no Argentinian Steam, and so on. Only idiots come up with this sorta policy. Brazil does the same shit as well. Would be cool if they stopped.

TA: You’ve worked in PC-98 and PSX aesthetics a few times. When .45 PARABELLUM BLOODHOUND was announced, I was floored by how it looked like the perfect game I’d want from you. The reception has been largely positive, but I can’t imagine what you and the team went through leading into its reveal with N1RV Ann-A and such. Can you talk about how the last few months have been for yourself?

CO: We’ve been locked in and doing our thing. No crunch, all fun. We party, we feast, we travel. We touch that proverbial grass. Lots of insecurities and self-doubt for sure. We even tried to downplay things and manage expectations before the reveal, bracing ourselves for possible apathy from the crowd because it’s not Nirvana. But when it came to actually putting in the work there was no hesitation. I’m happy the announcement worked out, but now we gotta buckle up and finish the story.

TA: .45 PARABELLUM BLOODHOUND has been revealed, people can wishlist it on Steam, and I immediately thought of Vagrant Story x Sukeban Games’ Vibe when I saw the trailer. How has it been interacting with fans discussing it online and offline?

CO: It’s been incredibly fun, even though there’s so many comparisons with old games. Not that I mind but there’s been some ridiculous comments out there that boggle the mind.

One thing that surprised me was the amount of fanart right after the reveal. That shit was great. A fan even brought us a drawing that we proudly displayed on our table at Bitsummit for everybody to see.

Our crown jewel

By @TumugiV !!! Thank you as always!!!!! pic.twitter.com/N1hbLMY25Q

— Sukeban Games (@SukebanGames) July 21, 2024

TA: Probably too soon but when can I buy the key art as a poster and get it signed?

CO: Maybe on release.

TA: What were your main inspirations for .45 PARABELLUM BLOODHOUND from a visual and gameplay perspective?

CO: When coming up with the actual gameplay for 45PB there was a unique concern. People know Sukeban as a Visual Novel/ADV developer because of Va11halla, and I wanted a sort of bridge between that audience and something more action-focused. I saw Parasite Eve’s battle system, but instead of saying “I want more of that” I thought “This is a good solution to my predicament” thanks to the hybrid of real time and turn-based gameplay. And that’s pretty much the way we work. Rather than wanting to imitate something, or wanting more of it, I look to games of the past in order to solve modern problems. We don’t need to reinvent the wheel every time.

For the visuals, I remember roaming the streets of Milan in 2019. I was ultra depressed, and didn’t actually have a country to live in at the time so I was also in a state of limbo. The mix of modern structures and old buildings, then juxtaposed to all the neon lights and LED screens along the river had my imagination running wild. For me that was the genesis for the game’s look.

This marriage between the old, the new and the decadent. The cherry on top was when I moved to Buenos Aires and found a similar atmosphere, except it included that South American roughness and texture that’s absent in Cyberpunk worlds (which tend to be mostly inspired by east asian aesthetics), so I decided to lean on that to create something unique.

TA: Tell us a little bit about the team working on it including the composer, and how long it has been in the works.

CO: We’re two people working on it day-to-day (me and the programmer), plus Merenge helping with additional character and production design. The composer this time around is Juneji, who has been doing a LOT of work with us over the years. We have this massive stash of incredible music that he’s made, but it remains unfortunately private due to circumstances with failed projects. Same with Merenge’s art.

There’s been some hella demoralizing moments since it’s hard to work on things for so long, pouring your soul into them only to see them crumble in front of you, but we’ve remained a tight group over the years and one of my goals is that the world can finally witness their talents on a large stage. I owe them that much. We also added an old friend in a producer/babysitting role and it’s helped us to not neglect important non-game stuff.

As for how long it’s been in development: Technically since 2019 but the current iteration is roughly 2 years old. Before that it was all experiments until we arrived at the right “moment to moment” gameplay.

TA: .45 PARABELLUM BLOODHOUND had a teaser, gameplay, and now has a steam page. Are there any plans to offer a demo on PC in any of Valve’s demo fests on Steam in the near future?

CO: It’d be a pain in the ass to maintain a demo for this particular game so we rather keep those for offline events. Never say never though.

TA: A lot of my friends who love VA-11 Hall-A are excited for .45 PARABELLUM BLOODHOUND. Will it be accessible for everyone or is it too soon to say how the difficulty will be handled?

CO: It’s too soon to explain, but as I’ve mentioned before, the battle system itself is meant to be a bridge between vibes-based players and action-oriented players. Not that I’m trying to satisfy both, but rather ease one type of player into a new system.

TA: What is your favorite aspect of .45 PARABELLUM BLOODHOUND right now?

CO: As a game? definitely the atmosphere and the script. I sometimes play it for leisure on my Steam Deck while tryna sleep and wonder “Damn, what’s gonna happen next?” only to be like “you wrote this shit! you know where it goes!”. Also the combat is fun as hell once it opens up after the first chapter. Chaining skills and upping the tempo of combat after pulling off sick tricks makes for addictive gameplay. This will make sense later unless we fumble it.

TA: Can you give us an interesting development / design anecdote for .45 PARABELLUM BLOODHOUND and VA-11 Hall-A?

CO: So I mentioned the influences of cities like Milan and Buenos Aires in the creation 45PB’s world, but one thing that people might have noticed by looking at very early screenshots was that it used to feature some locales reminiscent of Hong Kong. I actually scrapped a lot, if not all of it, in favor of the “South American Cyberpunk” I’m aiming for after a conversation I had with a friend from China.

I was consulting with him about getting the fonts and the text of the billboards just right so it felt authentic, but then my mind broke and I began asking to myself “Why go through all this effort trying to be authentic when I can just leverage my own culture?” and it’s a sort of mantra I’ve been following ever since. I find terms like “cultural appropriation” silly, but the word “appropriation” by itself is something I’ve been contending with a lot. “Is it my place to tell this type of story?” “Is it okay if these characters speak this or that way?” and so on. Not that it limits my creativity or that I’m afraid of hurting sensibilities; I frankly don’t give a shit. I just think that more original works can be born if we leverage what makes us unique instead of trying to make more of what we like. It’s all about balance.

TA: Since the announcement, you’ve no doubt had folks asking for a console version when we don’t even have a release year for .45 PARABELLUM BLOODHOUND, but I wanted to know if there was any thought into working with a publisher or will this be self-published?

CO: We want to self-publish on PC and let other companies handle consoles. We’re in the process of finding the right partner for this.

TA: What were the inspirations behind Reila Mikazuchi’s design and character?

CO: I don’t know if I’ll get in trouble for this but I admire actor and singer Meiko Kaji a lot.

I love her movies, you name them: Prisoner Scorpion, Stray Cat Rock, Lady Snowblood, Jeans Blues. Something in her look is so captivating and I wanted my own Meiko Kaji for 45PB. I needed a character design that can say a lot with just the eyes. Someone who can project that inner pain and tragedy effortlessly, so using Meiko Kaji as a point of reference was a must.

As for the writing, every time I come up with characters they’re always a composite of several people I know and myself. So as to not ruin your impressions of her, I shall keep these inspirations a secret.

TA: How many iterations did you go through for her final design which rules by the way?

CO: I always had in mind the main look of long, black hair and pale skin. Plus the third eye. The thing that took me the most work was coming up with the outfit. She was wearing a suit at first actually! Then I wanted her to have a jacket, but finding the right type of jacket was a challenge. I gave her a biker jacket but it looked weird and I had to experiment a lot until I stumbled upon the current design. Merenge helped me with some accessories, such as the shoes, gloves and the logo on the back of the jacket.

TA: VA-11 Hall-A saw VA-11 Hall-A Kids and Sapphic Pussy Rhapsody released following the game. Ahead of .45 PARABELLUM BLOODHOUND, should we expect any smaller projects like that?

CO: Never say never but our plan with this game is to release it, let it be and move on to something new. No DLC or anything like that. Ports to future platforms are always on the table but that’s the extent of it.

If A24 or someone else wants to make a movie about it I’ll listen.

TA: What does a day in your life look like right now?

CO: Right now I’m a little f***ed up. I’m usually a good boy who works from 9am, takes a little lunch break then locks in again until 4 or 5pm, but sleep escapes me lately so I’m just trying to contend with that in whatever way I can. The key is to not stress over not having much sleep. As long as shit gets done it’s all good. When I’m not working I try to catch a movie at one of the many small cinemas in my neck of the woods, or go outside and take a walk and buy books I’ll take forever to read.

Buenos Aires has this magic that inspires me to go outside and have low key adventures. Can’t get too crazy because this is still latin america and we don’t f*** around, but I love it here and being surrounded by so much culture, great food, and a crazy clubbing scene keeps me sane. Especially when there’s friends around. I owe them a lot for my mental health even if I’m a huge recluse sometimes. I’m the type to uninstall messaging apps for a minute when I really need to be alone.

TA: What have you been playing lately that you’ve enjoyed a lot?

CO: There’s so many video games. This year I really loved Children of the Sun and Arctic Eggs. Late last year I was really into The Citadel, Lethal Company and RoboCop: Rogue City. I’m now in the process of going through The Evil Within, which feels like a lost Grasshopper game sometimes. There’s the Elden Ring expansion, and I also recently replayed Kane and Lynch 2 with a friend and had a blast as always. That game is so f***ed up in the right ways. We need more grit like that.

TA: What do you think of the current state of indie games?

CO: Every time I go to a game event I’m always madly inspired by all the unique games I see at the indie spaces and that energy is partly what kept me going despite all our production hiccups. That sense of community. That we all have this burning desire to create for the sake of creation. That shit is great. I f***ing love indie games and I think they’re better now more than ever.

BUT… I worry that we are trying to lean too much on familiar concepts and inspirations.

Games like Arctic Eggs do it right in that it’s trying to look like an old PSX game, but there were no PSX games like Arctic Eggs. Then there’s games that feel so blatant in their desire to be like one of the classics that I’m like “weren’t we supposed to be the creative ones?” not to mention the over reliance on “Roguelike” mechanics and such. But it’s not all bad. There’s some really cool shit all the time; especially on places like itch.io. It’s just that humans in general can be creatively bankrupt regardless of production scale. There’s triple A slop, there’s indie slop. There’s no escaping the slop. Being indie doesn’t make you inherently more creative and that’s okay. There’s always room for growth as long as we’re cognizant about our shortcomings.

TA: Are you looking forward to playing any specific game this year?

CO: I’m really looking forward to Slitterhead.

Dunno if they’ll be out this year but there’s also Sonokuni, Elation For The Wonder Box 6000, Studio System: Guardian Angel, Eating Nature (from the Arctic Eggs dev)… There’s a lot I’m excited about. My Twitter feed is an endless stream of cool indie games I doubt will ever see the light of day but the attempt is appreciated nonetheless.

TA: Obviously scheduling and such, but if we managed doing this interview on a call, I’d have taken this chance to nerd out about The Silver Case with you. I’m glad I listened to my friend and powered through it because that typewriter sound was too much for me. I adore The Silver Case and I know you love it as well. What elements from it inspired you the most and what’s your favorite track from the soundtrack?

CO: The Silver Case has always been one of those games that felt like a white whale to me. It was so utterly inaccessible that my mind pretty much created its own Silver Case by trying to fill the gaps. I used to browse this creepy Suda51 fansite way back in the day and I was always searching for every bit of info on this inscrutable-ass game, so when it actually released in a language I can understand I was afraid I wasn’t gonna like it but I think it was much more than I could ever have imagined. In that sense, the way Silver Case inspires me is tightly related to this need to fill gaps. The space between the real Silver Case and what I was imagining it to be is where a lot of the ideas for the presentation in games like Va11halla or The Radio Wave Bureau were born.

I can’t choose a single favorite song if I’m being honest. The whole soundtrack has such a vibe. Could be the Sayaka Baian theme that plays in Kamuidrome, but tomorrow it could be the opening theme, then the next day another one.

TA: Did you end up trying it out on console or just playing it on PC?

CO: I bought and played the game on every platform.

TA: The Silver Case’s original box art and aesthetic is probably the only thing I love as much as that in VA-11 Hall-A. What elements from its visual style had you curious back then?

CO: Definitely the stoic look of every character. Takashi Miyamoto is my goat along with Yoshitoshi Abe. For Va11halla I had to be very expressive with the characters because that’s what the story commanded, but ever since I was a kid I always had an affinity towards designs that say a lot with minimal facial expressions. Two of my favorite films are Patlabor 2 and Ghost in the Shell, and Serial Experiments Lain is also one of my favorites, so you can probably tell where I’m coming from with this.

I was also so transfixed by the UI. I could go on and on about this subject, but I’ll always lament that The Silver Case didn’t create a movement after its release. Visual Novels could look so much better if we had 1% the imagination of that original Grasshopper team. I’m not sure, but I think the look was done by Akihiko Ishizaka who would go on to work on Killer7 and Danganronpa. I wish he was still working with GhM. Or maybe he still does and I’m talking out of my ass.

TA: I think you’ve met Suda more than once now. How has that been for you and has he played VA-11 Hall-A? I know Travis Strikes Again has a VA-11 Hall-A tshirt that is the objectively best and only one everyone should use.

CO: I’ve met the guy twice and I’ll always lament not knowing enough japanese to just talk and talk all day with him. We share a lot of common interests, though the generational gap can be felt when he starts yapping about Japanese media from the 70s. I do know he played my game but I’ve no clue if he actually enjoyed it. Will have to ask if I ever see him again.

TA: Are you still up for this if the opportunity arises?

CO: There’s a story behind this that I’ll have to sit on for a while longer.

TA: My current game of the year for 2024 is Like a Dragon: Infinite Wealth. I know you loved Yakuza: Like a Dragon, but did you end up trying Infinite Wealth/8 or Gaiden last year?

CO: I love Like a Dragon so much but I didn’t play Gaiden, and Infinite Wealth seemed like too much of a commitment when I played at launch. The game was throwing mechanic after mechanic at me in a way that made me feel overwhelmed. Like there was too much game ahead of me, so I put it off for the time being. Maybe once I’m in the mood again I’ll give it a honest shot.

TA: VA-11 Hall-A has always been perfect to play on a portable. I loved it on Switch and recently found a mod that adds full controller support to the PC version. I’ve been revisiting it on Steam Deck. Have you had a chance to try it on Steam Deck yet?

CO: I did and saw it sort of works but not quite. I wish I could go back and tweak it but it’s not technically feasible. The Game Maker version we used doesn’t even work properly on modern Windows and it’s the reason we haven’t been able to do any sort of fix or update to it. Aside from simply not having the time to deal with all that mess we made (at a technical level).

TA: Honestly I still have a lot more I want to discuss with you, but I think I should save that for part 2 or a future interview. Let me end this with one final question. How do you like your coffee? Go into as much detail as you can. If not coffee, talk about your favorite beverage.

CO: I like my coffee black like a moonless night. Even better if it’s accompanied by cheesecake on a beautiful afternoon.

TA: Probably not smart, but it is nearly 2 AM and sending these questions over made me want to replay The Silver Case. Next time we talk, let’s do a dedicated discussion on The Silver Case.

CO: Absolutely!

I’d like to thank Christopher Ortiz for their time and help with this interview over the last few weeks.

You can keep up with all our interviews here including our recent ones with FuturLab here, Shuhei Matsumoto from Capcom about Marvel Vs Capcom here, Santa Ragione here, Peter ‘Durante’ Thoman about PH3 and Falcom here, M2 discussing shmups and more here, Digital Extremes for Warframe mobile, Team NINJA, Sonic Dream Team, Hi-Fi Rush, Pentiment, and more. As usual, thanks for reading.

Source link

Heurist: A Comprehensive Overview

Web3

Metaverse Fashion Week Online

September 12, 2024

<![CDATA[

Heurist is a Layer 2 network designed for AI model hosting and inference, built on the ZK Stack. Its mission is to be the HuggingFace of Web3. By leveraging a decentralized computing resource network, Heurist provides serverless access to open-source AI models, utilizing blockchain technology to democratize AI and ensure accessibility and unbiased innovation.

Vision and Values

Heurist envisions AI to be instrumental in onboarding the next billion users to crypto. The project is grounded in the shared values of open-source AI and crypto, such as transparency, data ownership, and diversity. Heurist aims to foster a two-way onboarding process: empowering crypto enthusiasts to engage with AI through a permissionless mining network using their home PCs and enabling AI model developers to access increased compute resources, publicize their models, and earn revenue share.

By supporting the development of Web3-native AI applications, Heurist allows individuals to access AI services at low or no cost, showcasing the practical benefits of decentralized technologies. The Heurist ecosystem promotes innovation, collaboration, and inclusivity, uniting the strengths of AI and crypto to create a future where advanced technology is accessible to everyone.

Name and Inspiration

The name Heurist is derived from the concept of heuristics, which refers to mental shortcuts that help humans quickly arrive at reasonable solutions to complex problems, reflecting the projects goal to make sophisticated AI accessible and approachable for all users.

How Heurist was Founded

Jiewen Wang and Frank is the founder of Heurist. The journey began with an AI agent Heurist founders built to engage in crypto discussions on friend.tech. Through this project, they encountered several common challenges in AI Application development: high computational costs, censorship by major platforms like OpenAI, and a lack of AI model control.

To address these, they launched Heurist as a decentralized protocol to cut AI costs while maximizing customizability and sovereignty. With their teams extensive experience in Web2 and deep roots in crypto, the team is dedicated to democratizing technology and making an AI ecosystem safe, accessible, and resistant to central controls.

Heurist vision is to empower developers and users worldwide to harness the potential of AI without the limitations imposed by traditional centralized platforms.

The Heurist Team

Here’s how Heurist works🤔

Heurist protocol intricately connects various participants, each playing a vital role in maintaining a healthy decentralized ecosystem. These participants include Model Users (Consumers), Miners (Model Hosts), Model Creators, Application Integrators, and Validators.

To ensure economic security and the integrity of data returned from AI models, we will launch a utility token, Heurist Token (HUE). Participants can spend, stake and/or earn HUE based on the rules defined in smart contracts.

Roles in Heurist Ecosystem

Model Users: Users interact with the Heurist protocol to run inference tasks, such as text generation and image generation. They can choose from a variety of hosted AI models and pay for the computational resources on a pay-as-you-go basis.

Miners: Individuals with GPU resources can mine Heurist Token by hosting AI models. They run models on their own hardware and are compensated with user payments and Heurist Token emissions when consumers execute inference tasks. Miners are required to stake a certain amount of tokens as a commitment to providing quality services.

Model Creators: AI model creators drive the vibrancy of Heurist ecosystem. They upload AI models to the model registry of Heurist network and, in return, receive a portion of the payments made by users. This incentivizes creators to develop more advanced models to satisfy user’s growing demand.

Application Integrators: Application integrators provide interfaces to end users that use Heurists AI models under the hood. Such interfaces might include chatbots, AI agents, and image generation tools hosted as web applications, and might also include SDK that can be integrated in web services. Application integrators earn a fraction of fees in Heurist Token whenever consumers make a payment through the application.

Validators: Validators maintains the integrity and trustworthiness of the Heurist network. They periodically check the correctness of the data produced by miners. If a miner is found to be producing invalid or fake data, their stake is slashed, with a portion awarded to the validator who identified the discrepancy.

Dynamic Resource Allocation

As the network grows, an increasing number of AI models and compute resources (mainly GPUs) will join it. We use token voting to allocate compute resources to every model.

The heurist protocol allocates more powerful GPUs to those models with (1) higher demand and (2) high token voting. An AI model creator can acquire more compute resources in two ways: by encouraging organic usage or by voting with their own tokens. Token holders are incentivized to vote for higher-quality models with higher demand as they will receive a larger revenue share from those models.

The Current Project Update of Heurist

August has been a month of exciting launches, milestones, and community-driven initiatives. Lets dive into the highlights. You can refer to the blog Heurists August Highlights for full details.

zkImagine Launches on ZKsync: Heurist’s AI-powered NFT collection, zkImagine, allows users to create and mint unique artworks with zero gas fees, thanks to the Paymaster feature. Explore zkImagine.

1 Billion Inference Requests Processed: Heurist has processed over 1 billion inference requests, maintaining a strong performance with over 1,000 active GPUs. Mining emissions are stabilized at an annualized rate of 2.1%.

Heurist Heroes Game: A 50,000 HEU token grant has been approved for the Heurist Heroes Game, an engaging 2D endless shooter with AI quizzes and reward incentives.

%[https://vimeo.com/1005038193]

Create-to-Earn Campaign: Sprint 0 of the Create-to-Earn campaign is complete, refining a reward mechanism for users who create and mint NFTs with zkImagine.

Art Contest with Moni Talks: Over 100 images were submitted for the Russian-speaking communitys art contest, showcasing Heurist Imagines creative potential.

Heurist Roadmap

Heurist token (HUE)

HUE is a utility token with a dynamic supply, influenced by both emission and burning mechanisms. The maximum supply is capped at 1,000,000,000.

Emission Mechanisms

1. Mining

Process: Users can mine HEU tokens by hosting AI models on their GPUs.

Staking Requirement: A minimum of 10,000 esHEU tokens must be staked for a mining node to be active. Below this threshold, no rewards are generated.

Reward: Mining yields HEU tokens. Reward rate depends on GPU efficiency, availability (uptime), the type of AI model being run, and the total amount staked in a miner node.

Boosted POW Mining: For stakes between 10,000 to 100,000 HEU tokens, mining efficiency increases proportionally with the amount staked.

2. Staking

Process: Any user can stake esHEU tokens in mining nodes.

Reward: Staking yields are given in HEU token. Staking esHEU produces higher yields than HEU.

Vesting: esHEU rewards can be vested into HEU over a one-year period with linear vesting.

Transfer of Stake: Users can instantly transfer their HEU or esHEU stake from one miner to another. This promotes flexibility and competition among miners.

Burning Mechanisms

Similar to the EIP-1559 model in Ethereum, Heurist implements a token burn mechanism. When users pay for AI inference, a portion of the HEU payment is permanently removed from circulation.

The balance between token creation and burning is closely tied to network activity. During periods of high usage, the rate of token burning can exceed the rate of new token creation. This will potentially lead the Heurist network into a deflationary phase. This mechanism helps in regulating the token supply, and aligns the token value with the actual demand within the network.

Bribes and (3,3)

Bribes were a term first coined by users of DeFi, more specifically, by those of Curve Finance. These bribes are actually gamified incentivizes that help direct rewards for liquidity pools, in exchange for an incentive. We draw inspiration from the bribing mechanism of Curve, and apply it in boosted POW mining in Heurist.

Miners have the option to set a specific percentage of their mining rewards as a bribe to attract stakers. A staker may choose the miner with the highest bribe, but other factors should also affect the staker’s choice such as miner’s hardware performance and uptime.

Miners are incentivized to bribe because a higher stake in the miner node will result in higher mining efficiency. It creates a competitive yet cooperative environment where miners and stakers are aligned to provide better service to the network.

With this game theory-inspired (3,3) mechanism, we create an alignment between miners and token holders who take actions to achieve the best collective outcomes.

Proposal Page

Heurist Stats

Top 10 GPUs on Heurist Ecosystem

Heurist Unique Features and Advantages

Heurist offers many unique advantages and features:

Decentralized AI Network: Heurist is a decentralized protocol for AI model hosting and inference, built on the ZK Stack Layer 2 network. This decentralized approach aims to democratize AI and ensure accessibility, transparency, and unbiased innovation.

Incentivized Ecosystem: Heurist has a utility token (HUE) that facilitates an incentivized ecosystem. Participants can earn, stake, or spend HUE tokens based on their roles, such as miners (model hosts), model creators, application integrators, and validators.

Dynamic Resource Allocation: Heurist uses token voting to dynamically allocate compute resources (GPUs) to AI models based on demand and token voting. Models with higher demand and more token votes receive more powerful GPU resources.

Open-Source AI Models: Heurist emphasizes the use of open-source AI models, which are transparent, customizable, and often comparable in performance to closed-source models like ChatGPT or DALL-E.

Pay-as-you-go Model Access: Users can access and run inference tasks on a variety of hosted AI models, paying for computational resources on a pay-as-you-go basis, potentially at lower costs than closed-source alternatives.

Incentivized Testnet: Heurist has an incentivized testnet where participants can earn testnet rewards (Llama Points and Waifu Points) for hosting AI models on their GPUs. These points will be claimable as liquid HUE tokens at the mainnet launch.

Developer APIs: Heurist provides REST APIs for developers to integrate AI models, like Stable Diffusion for image generation and Mistral 8x7b (comparable to ChatGPT 3.5) for language models, into their applications.

Bribes and (3,3) Mechanism: Heurist implements a “bribes” mechanism inspired by DeFi, where miners can set a percentage of their mining rewards as a bribe to attract stakers, creating a competitive yet cooperative environment aligned with providing better services to the network.

Token Burning Mechanism: Heurist has a token burn mechanism similar to EIP-1559, where a portion of user payments for AI inference is permanently removed from circulation, potentially leading to a deflationary phase and aligning token value with network demand.

Anti-Cheat System: Heurist has an asynchronous monitoring system to track input and output of compute jobs, allowing for the slashing of testnet rewards for miners demonstrating malicious activities.

Node Rental Service: Heurist provides a managed GPU node rental service, sourcing deals from various data center partners, making it easier for participants without GPU hardware to join the network.

Community-Driven Innovation: Heurist aims to foster a culture of innovation, collaboration, and inclusivity, bringing together the AI and crypto communities to create an ecosystem where cutting-edge technology is accessible to all

Spheron X Heurist: Unleashing the Power of Decentralized AI

Heurist and Spheron are joining forces to redefine the landscape of AI and decentralized compute, empowering a community-driven future free from traditional barriers. By integrating Spherons cutting-edge, scalable compute solutions, Heurist will provide seamless access to open-source AI models for training and inference, significantly lowering costs and democratizing access to powerful AI tools.

In return, Spheron will leverage Heurists infrastructure to develop innovative, open-source use cases, showcasing the power of decentralized AI and GPU rental solutions. This partnership will also introduce a dedicated “GPU Rental” tab on Heurist’s platform, prominently featuring Spheron as a key partner, creating a unique synergy that accelerates the adoption of open, censorship-free AI technologies.

Heurist and Spheron are not just building tools; they are creating a decentralized, community-led ecosystem that prioritizes accessibility, transparency, and innovation in AI and compute.

Conclusion

In conclusion, Heurist is positioning itself at the forefront of a transformative movement in the AI and blockchain space. By leveraging decentralized computing and the ZK Stack Layer 2 network, Heurist aims to democratize AI access and foster innovation without the constraints of traditional centralized platforms.

The protocols unique structurewhere users, miners, developers, and validators collaborateensures a fair, incentivized, and transparent ecosystem that balances performance and community-driven governance. The vision of combining AI with crypto opens the door for billions to engage with advanced technologies in a cost-effective, unbiased, and censorship-free environment.

Through initiatives like zkImagine, the Heurist Heroes Game, and strategic partnerships with companies like Spheron, Heurist is advancing the technology and building a future where AI innovation is genuinely decentralized and accessible to everyone. As Heurist continues to grow, it is clear that the projects values of openness, collaboration, and inclusivity will be crucial to its long-term success.

FAQs

What is Heurist? Heurist is a decentralized Layer 2 network designed for AI model hosting and inference. It leverages blockchain technology to democratize AI and provide serverless access to open-source models.

How does Heurist work? Heurist connects various participantsmodel users, miners, creators, application integrators, and validatorsin a decentralized ecosystem where computational resources are allocated dynamically through token voting and incentives.

What role does the Heurist Token (HUE) play? HUE is the utility token of the Heurist network. It facilitates economic interactions like mining, staking, and payments for AI model usage, as well as governance through token voting.

What makes Heurist unique? Heurist offers decentralized AI services focusing on open-source models, pay-as-you-go access, a bribe mechanism for staking, and a token-burning system that can make the network deflationary during periods of high demand.

How does Heurist promote decentralized AI? Heurist decentralizes the hosting and usage of AI models, allowing individuals and developers to engage with AI services without relying on centralized platforms. Thus, it enhances accessibility, innovation, and fairness in the AI landscape.

]]>

Source link

GeForce NOW to Bring ‘Dead Rising Deluxe Remaster’ to the Cloud at Launch

Gaming

Metaverse Fashion Week Online

September 12, 2024

GeForce NOW to Bring ‘Dead Rising Deluxe Remaster’ to the Cloud at Launch

Rise and shine — Capcom’s latest action-adventure game, Dead Rising Deluxe Remaster, heads to the cloud at launch next week.

It’s part of nine new titles joining the extensive GeForce NOW library.

‘Dead Rising’ Coming Soon

From the ground to the cloud.

Dead Rising Deluxe Remaster returns with modern graphics. More than just a remaster, this Deluxe Remaster is a full graphical overhaul of the first game in the zombie-slaughtering action series Dead Rising. The remaster has also been fully voiced, supports auto-saves and has various other quality-of-life features.

One day, the peaceful town of Willamette, Colorado, found itself put under quarantine by the U.S. army. Frank West, a freelance journalist, smells a scoop and finds his way into the only shopping mall in town. Unfortunately, the mall has turned into a living hell, crawling with countless zombies. Help will arrive in 72 hours, so it’s up to him to find out the truth behind this incident before it’s too late.

Witness the unmatched mayhem and freedom when Dead Rising Deluxe Remaster launches on Wednesday, Sept. 19. Stream it with a GeForce NOW Priority or Ultimate membership for longer gaming sessions and higher frame rates.

New Games to Drive You Wild

Test Drive Unlimited Solar Crown, a new open-world racing game from KT Racing and Nacon, is now available for members to stream. Explore a fully recreated Hong Kong Island while taking to the road behind the wheels of exceptional cars and living the ultimate life of luxury. Test drive and purchase cars directly from dealerships, customize them in workshops and display them in the Solar Hotel garage. Each car offers a unique driving experience.

Members can look for the following games available to stream in the cloud this week:

Warhammer 40,000: Space Marine 2 (New release on Steam, Sept. 9)
Test Drive Unlimited Solar Crown (New release on Steam, Sept. 12)
Dawn of Defiance (Steam)
Flintlock: The Siege of Dawn (Xbox, available on PC Game Pass)
Fort Solis (Epic Games Store)
King Arthur: Legion IX (Steam)
Squirrel With a Gun (Steam)
Tyranny – Gold Edition (Xbox, available on Microsoft Store)
XIII (Xbox, available on Microsoft Store)

What are you planning to play this weekend? Let us know on X or in the comments below.

Which game(s) do you think would be cool as the opposite perspective? 🔀

— 🌩️ NVIDIA GeForce NOW (@NVIDIAGFN) September 10, 2024

Source link

Launchpad: The Future of Token Distribution

Web3

Metaverse Fashion Week Online

September 10, 2024

Launchpad: The Future of Token Distribution

We’re redefining how entire crypto ecosystems are built, launched, and scaled. Introducing the Layer3 Launchpad for exclusive token launches and ecosystem campaigns. This is tokenized attention at scale. Buckle up.

For explorers, Launchpad means:

Exclusive access to pre-token protocols

Incentives from established ecosystems

High-quality, interactive experiences

For projects, Launchpad is new rocket fuel:

Exposure to a massive, targeted user base

Onchain engagement plus flexible incentives

Token distribution with powerful Sybil filters

Launchpad isn’t just another platform. It’s the future of how crypto projects launch, grow, and align with their communities. We’re bridging the gap between emerging protocols and millions of engaged users, creating a new standard for token distribution and ecosystem expansion.

Launchpad on Layer3

Launching Now: Caldera 🌋🚀

Meet the Metalayer — the unifying layer of Ethereum rollups in our exclusive Launchpad campaign. Easily connect to new protocols via the Caldera Metalayer including Plume, B3, & Manta.

Featuring $100K in exclusive incentives for discovering Caldera’s rollup ecosystem.

Meet the Metalayer

L3 Token Staking

By staking L3, your loyalty unlocks expanded utility:

Boosted rewards on all campaigns

Early and extended access

Tiered benefits that scale with your stake

This is how you align with the future of crypto.

Source link

1...174175176...191 Page 175 of 191

Understanding GPU Memory Requirements for LLMs

Model Parameters (Weights)

How Model Size Impacts Memory

Key-Value (KV) Cache Memory

How Sequence Length and Concurrent Requests Impact KV Cache

Calculating KV Cache Memory

Activations and Temporary Buffers

Memory Overheads and Fragmentation

Calculating Total GPU Memory

Challenges in GPU Memory Optimization

Memory Optimization Techniques

PagedAttention: Reducing Memory Fragmentation with Paging

vLLM: A Near-Zero Memory Waste Solution

Managing Limited Memory

1. Swapping

2. Recomputation

Conclusion

FAQs

Understanding Behavioral Tokenomics

Understanding User Motivations

Incentives and Rewards

Social Proof and Network Effects

Token Utility and Behavioral Levers

Understanding Behavioral Tokenomics

Intersection of Behavioral Economics and Tokenomics

Influence of Psychological Factors on Token Attraction

The Effect of Presentation and Context

Implications for Token Designers

Leveraging Price Performance

Scarcity and Perceived Value

Initial Supply Considerations

Managing Maximum Supply and Inflation

Loss Aversion

Social Proof

Mental Accounting

Endowment Effect

Conclusion

FAQ

Examples of Virtual Holiday Stores

J.Crew Virtual Ski Chalet

Corona Holiday Beach Hut

Sephora X Cosmopolitan Gifting Suite

H-E-B Grocery Holiday Home

Laneige Virtual World

What Is Advanced Reasoning in AI?

Why Is the Model Called “Strawberry”?

Key Features of GPT-01 “Strawberry”

1. Coding: Thoroughly Analyzing Each Instruction

2. Reasoning: Understanding Context and Surroundings

3. Mathematics: Tackling Complex Problems

Comparing GPT-01 to GPT-4

Use Cases for GPT-01

Coding and Programming

Mathematical Problem Solving

Business Applications

Education and Tutoring

Challenges with GPT-01

What’s Next for OpenAI?

Conclusion

FAQs

1. Start with Skincare

2. Prime for Perfection

3. Keep It Light with Foundation

4. Conceal Strategically

5. Subtle Definition with Brows

6. Enhance Your Eyes Softly

7. Add a Natural Flush

8. Luminous Lips

9. Set with a Mist

The Rise of Collaborative Learning in AI

The Role of Synthetic Data in Transfer Learning

Self-Instruct: Enhancing Models with Self-Generated Tasks

Alpaca: Efficient AI with Teacher-Student Dynamics

How Collaborative Learning is Democratizing AI

The Future of Collaborative AI

Learning Paradigms: Beyond Traditional Fine-Tuning

1. Iterative Refinement: Self-Improvement Through Continuous Feedback

2. Knowledge Distillation: Transferring Wisdom from Teacher to Student

3. Learning by Teaching: A Two-Way Feedback Loop for Continuous Improvement

Philosophical Implications: Machines Learning from Machines