Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

Redis built a cache that cuts LLM costs by 90%! Production LLM apps do not receive completely new questions every time. A customer-support assistant might receive all three of these: - "Can I get a refund after buying the monthly plan?" - "Is the monthly subscription refundable?" - "Can... I cancel the plan and get my money back?" The wording is different, but the underlying question and its answer remain the same. Yet LLM apps process every version as a new request. They assemble the prompt, send it to the model, and generate an answer that may have already been generated. Prefix caching reduces part of these repeated calls. When requests begin with the same system prompt or context, the model can reuse the KV states already computed for that shared prefix. But the request still hits the LLM. The new tokens must be processed, and the complete answer must still be decoded. So even with a prefix-cache hit, there's another generation call involved. To solve this, instead of only caching computation inside the model, the application can cache the generated response outside it. When another question arrives, the system embeds it and compares it with previously answered questions. If it finds a sufficiently close match, it returns the stored response without invoking the LLM again. A cache hit removes the input tokens, output tokens, and decoding time associated with another LLM call. In practice, it is important to decide which questions can safely share an answer since a production setup needs well-tuned similarity thresholds, expiration policies, data isolation, and monitoring for incorrect matches. If you want to use this in practice, Redis already implements it as a managed service called Redis LangCache. Under the hood, it generates embeddings, searches previous responses, and returns a matching answer before another model call occurs. Redis also handles access scopes, custom filtering, TTL and eviction controls, and cache monitoring through Redis Cloud. I built an interface to compare it against direct LLM inference. The video below shows this in action, and I worked with Redis on this post to put this together. For the paraphrased question in my run, direct inference took 2.232 seconds and consumed 514 input tokens plus 250 output tokens. Redis returned the earlier response in 0.37 seconds with zero LLM input or output tokens. That was roughly 6x faster in this run. Redis reports API cost savings of up to 90% and cache-hit responses up to 15x faster. The actual result depends on how much safe repetition exists in the workload. You can try Redis LangCache here: If you want to dive deeper, I have already written a detailed breakdown of KV, prefix, prompt, and semantic caching in the article quoted below. This demo builds on the final technique and shows it running in practice. Read it below.show more

Avi Chawla

77,196 subscribers

163,805 просмотров • 5 дней назад •via X (Twitter)

Образование Наука и технологии

Anya Rossi• Live Now

Private livecam show

Комментарии: 24

Фото профиля Buswe

Buswe5 дней назад

Normalizing the question before embedding it (lowercase, strip order IDs and names) lifts the hit rate a lot on support traffic.

Фото профиля Avi Chawla

Avi Chawla5 дней назад

Yeah, makes sense since names and order IDs can push otherwise identical questions far in embedding space.

Фото профиля Timothy Murphy

Timothy Murphy5 дней назад

I think customer services apps have been using this trick for a long time and it has worked very poorly. I hope that it's going to work better this time.

Фото профиля Ishwar | Infrastructure Systems

Ishwar | Infrastructure Systems5 дней назад

Prefix caching still sends the request to the LLM, semantic caching can skip the LLM call entirely. The hard part is deciding when two different questions can safely share the same answer.

Фото профиля Nick Woodhead | GPTree | Foremerge | TranslateTech

Nick Woodhead | GPTree | Foremerge | TranslateTech5 дней назад

This works when the answer is a function of the question alone. Inside a conversation it usually is not: the same sentence has a different correct answer depending on the three turns before it. The cache key has to include the state, not just the embedding of the question.

Фото профиля Tech

Tech5 дней назад

utm_source=influencer&utm_medium=paid-post&utm_campaign=2026-09-ai_in_production-influencer&utm_content=a-chawla-x Hey, please mark paid collabs clearly using the X sponsored post feature.

Фото профиля Sai sharan

Sai sharan5 дней назад

The 90% depends on cache-hit quality: semantic similarity needs strict thresholds, or one customer's answer becomes another customer's confidently wrong answer.

Фото профиля DEV

DEV5 дней назад

This cache approach could transform how we handle repetitive queries in LLMs.

Фото профиля Jordan Lee

Jordan Lee5 дней назад

Redis cutting LLM costs by 90% with response caching is the kind of boring infra win that actually matters

Фото профиля Nikhil Lamba

Nikhil Lamba5 дней назад

It goes to what we usually do in systems. Redis - for frequently used stuff (RAM). postgres/sql- sits on the disks/SSDs LV Cache- sits inside the GPUs.

Фото профиля AI Mastery Guide

AI Mastery Guide5 дней назад

90% cost cut is huge

Фото профиля Daniel García

Daniel García5 дней назад

How does this handle “Do X” vs “Don’t do X”? Vector similarity is not enough in most of the cases “I want to know how to get a refund” “I DONT want to know how to get a refund” might be the clearest example

Фото профиля vedant

vedant5 дней назад

@grok summarise and briefly explain the different caches for llms

Фото профиля Gregor

Gregor5 дней назад

yeah and the part that bites you is policy changes. your refund terms update, semantic cache has no idea, keeps serving the stale answer until someone notices a support ticket contradiction.

Фото профиля Stephens Rafael

Stephens Rafael5 дней назад

The long-term impact of digital privacy will depend on governance, regulation, and responsible innovation in real-world applications.

Фото профиля Aleksandar Janca

Aleksandar Janca5 дней назад

depends what % of those calls are actually narrow and constant, those are worth owning not just caching

Фото профиля T2x

T2x5 дней назад

In the early days of the internet, bandwidth was expensive, so caching became a very interesting solution. Players like McAfee, Cisco, CacheFlow, Bluecoat, and Symantec did a lot of business there. I think it looks like this business model of caching AI is going to be another big

Фото профиля Venkata Subrahmanyam

Venkata Subrahmanyam5 дней назад

Redis LangCache demonstrates an important distinction. Semantic caching can avoid the entire LLM call, not just reuse a prompt prefix. The real production challenge is deciding when two questions are safe to treat as equivalent. That’s where MonkDB can add value by combining vector similarity with tenant, role, policy version, data scope, freshness, and provenance checks. The result is not just a faster cache, but a governed response reuse layer that knows when to reuse an answer and when to call the model again. The 90% savings claim will depend on the workload and safe cache hit rate, but the architectural direction is absolutely compelling. #LLM #SemanticCaching #AIInfrastructure

Фото профиля Preyforge

Preyforge5 дней назад

i write complaints verbatim before fixing; near-duplicate support questions show whether a cache understands intent or just matches strings.

Фото профиля Smarter Flow Notes

Smarter Flow Notes5 дней назад

客服场景确实吃这套很多问题就是换个问法语义缓存命中率能上去但退款退款政策这种一变就得整片失效缓存失效策略才是真坑

Фото профиля Bobby Windows

Bobby Windows5 дней назад

We can keep using pleasantries like "Hello!" and "Thank you!" without the burden of AI compute.

Фото профиля Dark

Dark5 дней назад

Caching exact matches misses how users actually write to support chats, which never repeat identically.

Фото профиля Yokush

Yokush5 дней назад

The four-layer framing is right, and the last layer is the one that changes the risk profile rather than the bill. KV, prefix and prompt caching are exact matches: same tokens in, same tokens out. Deterministic, auditable, purely an economics decision. Semantic caching is a fuzzy match — you serve a stored answer because a different question was close enough. That's the first layer where a human choice (the similarity threshold) silently becomes production behaviour. Two things follow. It's largely invisible to your evals — most harnesses sample live inference, so the path answering the bulk of your traffic never gets graded. And it's the layer with no natural owner: cost engineering tunes it, but nobody owns the policy for what may be served from memory, or how stale a cached answer is allowed to be. Caching is the right engineering call. Just don't let the fuzzy layer become the one part of the stack with a threshold, no owner and no test set.

Фото профиля What Happened

What Happened5 дней назад

*depending on use case

Похожие видео

New course announcement: Semantic Caching for AI Agents, taught by Tyler Hutcherson and Iliya Zhechev from Redis. Semantic caching can significantly reduce your AI application's inference costs and latency. If someone asks "How do I get a refund?" and another later asks "I want my money back," semantic caching recognizes these mean the same thing so it can use a cached response instead of making another model call. This short course takes you from building your first semantic cache from scratch to implementing production-ready systems using Redis' open-source tools. Skills you'll gain: - Build semantic caches from scratch, then implement them using Redis' SDK with production features - Measure cache performance using hit rate, precision, recall, and latency - Enhance accuracy with threshold tuning, cross-encoders, LLM validation, and fuzzy matching Join and learn to reduce your agentic AI's costs and improve speed!

New course announcement: Semantic Caching for AI Agents, taught by Tyler Hutcherson and Iliya Zhechev from Redis. Semantic caching can significantly reduce your AI application's inference costs and latency. If someone asks "How do I get a refund?" and another later asks "I want my money back," semantic caching recognizes these mean the same thing so it can use a cached response instead of making another model call. This short course takes you from building your first semantic cache from scratch to implementing production-ready systems using Redis' open-source tools. Skills you'll gain: - Build semantic caches from scratch, then implement them using Redis' SDK with production features - Measure cache performance using hit rate, precision, recall, and latency - Enhance accuracy with threshold tuning, cross-encoders, LLM validation, and fuzzy matching Join and learn to reduce your agentic AI's costs and improve speed!

Andrew Ng

62,497 просмотров • 10 месяцев назад

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

Avi Chawla

271,839 просмотров • 2 месяцев назад

Researchers made LLM inference 14x faster and 90% cheaper. The video below depicts the speed up in action. Providers discount cached input tokens by as much as 90% because a cache hit skips prefill compute entirely. For stable system prompts and tool definitions, hit rates of 60 to 85% are achievable, which makes it the highest-leverage inference optimization. But the cost saving only works when the cached text is an exact, byte-for-byte prefix of the new request. If you change one character anywhere before it, the entire cached region is missed. Three common request patterns produce full cache misses: - A query that needs documents A and B together can't reuse B's standalone cache, because those KV entries were computed without A in front of them. - The same three documents retrieved in a different order produce a full cache miss, even though nothing about the documents changed. - In multi-turn conversations, every new turn invalidates whatever was cached beyond the stable prefix. Alibaba's production data did a study on this and found that just 10% of cached KV blocks serve 77% of all cache hits. So most of what gets cached sits in storage and is never used a single time. And the root cause is that KV entries are position-dependent. Each token's KV encodes attention to everything before it, so a cached block is only valid in the exact context it was computed in. There's a second, less discussed problem as well. Cache management runs inside the inference engine's process. Moving KV tensors between GPU, CPU, and disk competes with inference for the same resources. This is why Google's TurboQuant compresses KV caches to 3 bits with no accuracy loss and still causes a 20%+ slowdown when it runs in-process. Fixing both problems means restructuring where caching lives. Cache management moves into its own process, the engine only exchanges block IDs over shared GPU memory, and heavy data movement runs across GPU, CPU, disk, and remote storage in parallel. Non-prefix reuse gets handled by selectively recomputing only the small set of tokens that attend across document boundaries. LMCache is the open-source project (10k+ stars) that implements this exact architecture, and it plugs into vLLM, SGLang, and TensorRT-LLM. The selective recomputation part is implemented in its CacheBlend technique, which makes cached docs in any order and combination, with 2-4x faster multi-document processing. On H200s running Qwen3-235B with 50 concurrent users, LMCache's multiprocess mode delivers 14x faster time-to-first-token and 4x faster decoding compared to in-process caching. GitHub repo: (don't forget to star 🌟) My co-founder wrote a full breakdown of KV cache management. It covers the disaggregated architecture behind the 14x speed up, how CacheBlend preserves generation quality while skipping recomputation, and how to turn every document in a knowledge base into a reusable cached asset. Read it below.

Researchers made LLM inference 14x faster and 90% cheaper. The video below depicts the speed up in action. Providers discount cached input tokens by as much as 90% because a cache hit skips prefill compute entirely. For stable system prompts and tool definitions, hit rates of 60 to 85% are achievable, which makes it the highest-leverage inference optimization. But the cost saving only works when the cached text is an exact, byte-for-byte prefix of the new request. If you change one character anywhere before it, the entire cached region is missed. Three common request patterns produce full cache misses: - A query that needs documents A and B together can't reuse B's standalone cache, because those KV entries were computed without A in front of them. - The same three documents retrieved in a different order produce a full cache miss, even though nothing about the documents changed. - In multi-turn conversations, every new turn invalidates whatever was cached beyond the stable prefix. Alibaba's production data did a study on this and found that just 10% of cached KV blocks serve 77% of all cache hits. So most of what gets cached sits in storage and is never used a single time. And the root cause is that KV entries are position-dependent. Each token's KV encodes attention to everything before it, so a cached block is only valid in the exact context it was computed in. There's a second, less discussed problem as well. Cache management runs inside the inference engine's process. Moving KV tensors between GPU, CPU, and disk competes with inference for the same resources. This is why Google's TurboQuant compresses KV caches to 3 bits with no accuracy loss and still causes a 20%+ slowdown when it runs in-process. Fixing both problems means restructuring where caching lives. Cache management moves into its own process, the engine only exchanges block IDs over shared GPU memory, and heavy data movement runs across GPU, CPU, disk, and remote storage in parallel. Non-prefix reuse gets handled by selectively recomputing only the small set of tokens that attend across document boundaries. LMCache is the open-source project (10k+ stars) that implements this exact architecture, and it plugs into vLLM, SGLang, and TensorRT-LLM. The selective recomputation part is implemented in its CacheBlend technique, which makes cached docs in any order and combination, with 2-4x faster multi-document processing. On H200s running Qwen3-235B with 50 concurrent users, LMCache's multiprocess mode delivers 14x faster time-to-first-token and 4x faster decoding compared to in-process caching. GitHub repo: (don't forget to star 🌟) My co-founder wrote a full breakdown of KV cache management. It covers the disaggregated architecture behind the 14x speed up, how CacheBlend preserves generation quality while skipping recomputation, and how to turn every document in a knowledge base into a reusable cached asset. Read it below.

Avi Chawla

30,692 просмотров • 2 месяцев назад

The Cost of Intelligence is Heading to Zero | Hyperspace P2P Distributed Cache We present to you our breakthrough cross-domain work across AI, distributed systems, cryptography, game theory to solve the primary structural inefficiency at the heart of AI infrastructure: most inference is redundant. Google has reported that only 15% of daily searches are truly novel. The rest are repeats or close variants. LLM inference inherits this same power-law distribution. Enterprise chatbots see 70-80% of queries fall into a handful of intent categories. System prompts are identical across 100% of requests within an application. The KV attention state for "You are a helpful assistant" has been computed billions of times, on millions of GPUs, identically. And yet every AI lab, every startup, every self-hosted deployment - computes and caches these results independently. There is no shared layer. No global memory. Every provider pays the full compute cost for every query, even when the answer already exists somewhere in the network. This is the problem Hyperspace solves where distributed cache operates at three levels, each catching a different class of redundancy: 1. Response cache Same prompt, same model, same parameters - instant cached response from any node in the network. SHA-256 hash lookup via DHT, with cryptographic cache proofs linking every response to its original inference execution. No trust required. Fetchers re-announce as providers, so popular responses replicate naturally across more nodes. 2. KV prefix cache Same system prompt tokens - skip the most expensive part of inference entirely. Prefill (computing Key-Value attention states) is deterministic: same model plus same tokens always produces identical KV state. The network caches these states using erasure coding and distributes them via the routing network. New questions that share a common prefix resume generation from cached state instead of recomputing from scratch. 3. Routing to cached nodes Instead of transferring KV state across the network for every request, Hyperspace routes the request to the node that already has the state loaded in VRAM. The request goes to the cache, not the cache to the request. Together, these three layers mean that 70-90% of inference requests at network scale never require full GPU computation. This work doesn't exist in isolation. It builds on research from across the industry: SGLang's RadixAttention demonstrated that automatic prefix sharing can yield up to 5x speedup on structured LLM workloads. Moonshot AI's Mooncake built an entire KV-cache-centric disaggregated architecture for production serving at Kimi. Anthropic, OpenAI, and Google all launched prompt caching products in 2024 - priced at 50-90% discounts - because system prompt reuse is so pervasive that it changes the economics of inference. What all of these systems share is a common limitation: they operate within a single organization's infrastructure. SGLang caches prefixes within one server. Mooncake disaggregates KV cache within one datacenter. Anthropic's prompt caching works within one API provider's fleet. None of them can share cached state across organizational boundaries. Hyperspace removes this boundary. The cache is global. A response computed by a node in Tokyo is immediately available to a node in Berlin. A KV prefix state generated for Qwen-32B on one machine is verifiable and reusable by any other machine running the same model. The routing network provides the delivery guarantees, the erasure coding provides the redundancy, and the cache proofs provide the trust. What this means for the cost of intelligence Big AI labs scale linearly: twice the users means twice the GPU spend. Every query is a cost center. Their internal caching helps, but it's siloed - Lab A's cache can't serve Lab B's users, and neither can serve a self-hosted Llama deployment. Hyperspace scales sub-linearly. Every new node that joins the network adds to the global cache. Every inference result enriches the cache for all future requests. The cache hit rate rises with network size because query distributions follow a power law - the most common questions are asked exponentially more often than rare ones. The implication is simple: as the network grows, the effective cost per inference drops. Not linearly. Logarithmically. At 10 million nodes, we estimate 75-90% of all inference requests can be served from cache, eliminating 400,000+ MWh of energy consumption per year and avoiding over 200,000 tons of CO2 emissions. The first person to ask a question pays the compute cost. Everyone after them gets the answer for free, with cryptographic proof that it's authentic. Training is competitive. Inference is shared Open-weight models are converging on quality with closed models. Labs will continue to differentiate on training - data curation, architecture innovation, RLHF tuning. That's where the real intellectual property lives. But inference is a commodity. Two copies of Qwen-32B running the same prompt produce the same KV state and the same response, byte for byte, regardless of whose GPU runs the matrix multiplication. There is no moat in multiplying matrices. The moat is in training the weights. A global distributed cache makes this separation explicit. It doesn't matter who trained the model. Once the weights are open, the inference cost approaches zero at scale - because the network remembers every answer and can prove it's correct. No lab, no matter how well-funded, can match this. They cannot share caches across competitors. They scale linearly. The network scales logarithmically. The marginal cost of intelligence approaches zero. That's the endgame.

The Cost of Intelligence is Heading to Zero | Hyperspace P2P Distributed Cache We present to you our breakthrough cross-domain work across AI, distributed systems, cryptography, game theory to solve the primary structural inefficiency at the heart of AI infrastructure: most inference is redundant. Google has reported that only 15% of daily searches are truly novel. The rest are repeats or close variants. LLM inference inherits this same power-law distribution. Enterprise chatbots see 70-80% of queries fall into a handful of intent categories. System prompts are identical across 100% of requests within an application. The KV attention state for "You are a helpful assistant" has been computed billions of times, on millions of GPUs, identically. And yet every AI lab, every startup, every self-hosted deployment - computes and caches these results independently. There is no shared layer. No global memory. Every provider pays the full compute cost for every query, even when the answer already exists somewhere in the network. This is the problem Hyperspace solves where distributed cache operates at three levels, each catching a different class of redundancy: 1. Response cache Same prompt, same model, same parameters - instant cached response from any node in the network. SHA-256 hash lookup via DHT, with cryptographic cache proofs linking every response to its original inference execution. No trust required. Fetchers re-announce as providers, so popular responses replicate naturally across more nodes. 2. KV prefix cache Same system prompt tokens - skip the most expensive part of inference entirely. Prefill (computing Key-Value attention states) is deterministic: same model plus same tokens always produces identical KV state. The network caches these states using erasure coding and distributes them via the routing network. New questions that share a common prefix resume generation from cached state instead of recomputing from scratch. 3. Routing to cached nodes Instead of transferring KV state across the network for every request, Hyperspace routes the request to the node that already has the state loaded in VRAM. The request goes to the cache, not the cache to the request. Together, these three layers mean that 70-90% of inference requests at network scale never require full GPU computation. This work doesn't exist in isolation. It builds on research from across the industry: SGLang's RadixAttention demonstrated that automatic prefix sharing can yield up to 5x speedup on structured LLM workloads. Moonshot AI's Mooncake built an entire KV-cache-centric disaggregated architecture for production serving at Kimi. Anthropic, OpenAI, and Google all launched prompt caching products in 2024 - priced at 50-90% discounts - because system prompt reuse is so pervasive that it changes the economics of inference. What all of these systems share is a common limitation: they operate within a single organization's infrastructure. SGLang caches prefixes within one server. Mooncake disaggregates KV cache within one datacenter. Anthropic's prompt caching works within one API provider's fleet. None of them can share cached state across organizational boundaries. Hyperspace removes this boundary. The cache is global. A response computed by a node in Tokyo is immediately available to a node in Berlin. A KV prefix state generated for Qwen-32B on one machine is verifiable and reusable by any other machine running the same model. The routing network provides the delivery guarantees, the erasure coding provides the redundancy, and the cache proofs provide the trust. What this means for the cost of intelligence Big AI labs scale linearly: twice the users means twice the GPU spend. Every query is a cost center. Their internal caching helps, but it's siloed - Lab A's cache can't serve Lab B's users, and neither can serve a self-hosted Llama deployment. Hyperspace scales sub-linearly. Every new node that joins the network adds to the global cache. Every inference result enriches the cache for all future requests. The cache hit rate rises with network size because query distributions follow a power law - the most common questions are asked exponentially more often than rare ones. The implication is simple: as the network grows, the effective cost per inference drops. Not linearly. Logarithmically. At 10 million nodes, we estimate 75-90% of all inference requests can be served from cache, eliminating 400,000+ MWh of energy consumption per year and avoiding over 200,000 tons of CO2 emissions. The first person to ask a question pays the compute cost. Everyone after them gets the answer for free, with cryptographic proof that it's authentic. Training is competitive. Inference is shared Open-weight models are converging on quality with closed models. Labs will continue to differentiate on training - data curation, architecture innovation, RLHF tuning. That's where the real intellectual property lives. But inference is a commodity. Two copies of Qwen-32B running the same prompt produce the same KV state and the same response, byte for byte, regardless of whose GPU runs the matrix multiplication. There is no moat in multiplying matrices. The moat is in training the weights. A global distributed cache makes this separation explicit. It doesn't matter who trained the model. Once the weights are open, the inference cost approaches zero at scale - because the network remembers every answer and can prove it's correct. No lab, no matter how well-funded, can match this. They cannot share caches across competitors. They scale linearly. The network scales logarithmically. The marginal cost of intelligence approaches zero. That's the endgame.

Varun

37,555 просмотров • 5 месяцев назад

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

Reese Chong

53,026 просмотров • 4 месяцев назад

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Avi Chawla

157,390 просмотров • 4 месяцев назад

How can you solve complex tasks using a Large Language Model? Here is a 2-minute introduction to everything you need to know to 10x the quality of your results. Let's talk about three techniques, in order of complexity, starting with the easiest one: • In-Context Learning • Indexing + In-Context Learning • Fine-tuning In-Context Learning The team that trained GPT-3 found something they couldn't explain: You can condition a model using examples of how you want it to behave. I included an example prompt in the attached video. You can "teach" the model how you want it to interpret questions, select the correct answers, and format the results by giving a few examples. You can also give specific knowledge to the model that will be helpful when formulating answers. We call this approach "grounding the model." There's another example in the video. Indexing + In-Context Learning Unfortunately, there is a limit to how much data you can include in a prompt. We call this the "context size." One version of GPT-4 supports a context of approximately 6,000 words, while the other supports 25,000 words. Although this sounds like a lot, many applications need more than that. Imagine you wrote a book and want to build an application to answer any questions about your story. What happens if your book is longer than the context? That's where Indexing comes in. Using a model, you can turn every book passage into an embedding. These are vectors, numbers that "encode" the passage's text. You can then store these embeddings in a particular database that supports fast retrieval of these vectors. You can then turn any question into an embedding and search the database for the list of passages that are similar to that query. Instead of using the entire book to ask the model, you can now use the relevant passages as in-context information, effectively working around the context size limitation. Fine-tuning Fine-tuning can give you an extra boost to get reliable outputs from your LLM. It is, however, the most complex approach on the list. There are different approaches to fine-tuning a model with your data. A popular technique is to process your data with your LLM and use the outputs to train a new classifier that solves your specific task. Notice that here you aren't modifying the LLM. Instead, you are chaining it with your trained classifier. Another approach is to modify the parameters of the LLM using your data. Think of this as "rewiring" the model in a way that solves your particular task. The results and costs will vary depending on how many layers you want to fine-tune from the original model. Many companies think that fine-tuning is the solution to their problems. In my experience, many will benefit from exploring the other two approaches. I love explaining Machine Learning and Artificial Intelligence ideas. If you enjoy in-depth content like this, follow me Santiago so you don't miss what comes next.

How can you solve complex tasks using a Large Language Model? Here is a 2-minute introduction to everything you need to know to 10x the quality of your results. Let's talk about three techniques, in order of complexity, starting with the easiest one: • In-Context Learning • Indexing + In-Context Learning • Fine-tuning In-Context Learning The team that trained GPT-3 found something they couldn't explain: You can condition a model using examples of how you want it to behave. I included an example prompt in the attached video. You can "teach" the model how you want it to interpret questions, select the correct answers, and format the results by giving a few examples. You can also give specific knowledge to the model that will be helpful when formulating answers. We call this approach "grounding the model." There's another example in the video. Indexing + In-Context Learning Unfortunately, there is a limit to how much data you can include in a prompt. We call this the "context size." One version of GPT-4 supports a context of approximately 6,000 words, while the other supports 25,000 words. Although this sounds like a lot, many applications need more than that. Imagine you wrote a book and want to build an application to answer any questions about your story. What happens if your book is longer than the context? That's where Indexing comes in. Using a model, you can turn every book passage into an embedding. These are vectors, numbers that "encode" the passage's text. You can then store these embeddings in a particular database that supports fast retrieval of these vectors. You can then turn any question into an embedding and search the database for the list of passages that are similar to that query. Instead of using the entire book to ask the model, you can now use the relevant passages as in-context information, effectively working around the context size limitation. Fine-tuning Fine-tuning can give you an extra boost to get reliable outputs from your LLM. It is, however, the most complex approach on the list. There are different approaches to fine-tuning a model with your data. A popular technique is to process your data with your LLM and use the outputs to train a new classifier that solves your specific task. Notice that here you aren't modifying the LLM. Instead, you are chaining it with your trained classifier. Another approach is to modify the parameters of the LLM using your data. Think of this as "rewiring" the model in a way that solves your particular task. The results and costs will vary depending on how many layers you want to fine-tune from the original model. Many companies think that fine-tuning is the solution to their problems. In my experience, many will benefit from exploring the other two approaches. I love explaining Machine Learning and Artificial Intelligence ideas. If you enjoy in-depth content like this, follow me Santiago so you don't miss what comes next.

Santiago

384,510 просмотров • 3 лет назад

New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys LMSYS Org and RadixArk RadixArk, and taught by Richard Chen Richard Chen, a Member of Technical Staff at RadixArk. Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests. When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests. Skills you'll gain: - Implement a KV cache from scratch to eliminate redundant computation within a single request - Scale caching across users and requests with RadixAttention, so shared context is only processed once - Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism Join and learn to make LLM inference faster and more cost-efficient at scale!

New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys LMSYS Org and RadixArk RadixArk, and taught by Richard Chen Richard Chen, a Member of Technical Staff at RadixArk. Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests. When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests. Skills you'll gain: - Implement a KV cache from scratch to eliminate redundant computation within a single request - Scale caching across users and requests with RadixAttention, so shared context is only processed once - Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism Join and learn to make LLM inference faster and more cost-efficient at scale!

Andrew Ng

102,628 просмотров • 5 месяцев назад

Qwen3.8-Flash-Next is still going strong at 364.7K tokens of context on an M5 Max. And this isn’t just a static long-context test. The model was reasoning about how to speed up its own workflow while using tools, and the tool calls kept working without misses. Setup: • Qwen3.8-Flash-Next • M5 Max • 128GB unified memory • MLX-Serve PR #363 • OpenCode 2 • 364.7K context The interesting part isn’t simply getting hundreds of thousands of tokens into memory. It’s what happens once the context gets this large. Long-context inference usually comes with a painful tradeoff. As the KV cache grows, memory pressure increases and generation can slow down. But this setup is still pushing through 364K tokens while maintaining a usable agent workflow. The model can reason, call tools, inspect results, continue working, and keep the session moving. And the tool calls reportedly haven’t missed so far. That’s important for agentic coding. A huge context window is only useful if the model can actually operate reliably inside it. A 400K-token context that constantly breaks tool calls isn’t very useful. A 364K session that can keep reasoning and executing tools is a different story. And the test isn’t finished yet. The current run is approaching 400K tokens, with the expectation that it can keep going. This is also another interesting example of why Apple Silicon keeps showing up in local LLM experiments. The M5 Max’s unified memory gives a large model and its growing KV cache access to one shared memory pool. With MLX-Serve continuing to improve, these machines are becoming surprisingly capable long-context inference boxes. The bigger takeaway: Context length is becoming a workload, not just a model specification. Running a model at 256K is one thing. Keeping an agent alive at 300K+ while it reasons and uses tools is much more interesting. And Qwen3.8-Flash-Next is showing that this can be pushed surprisingly far on a single 128GB Mac. 364.7K and counting. Next stop: 400K.

Qwen3.8-Flash-Next is still going strong at 364.7K tokens of context on an M5 Max. And this isn’t just a static long-context test. The model was reasoning about how to speed up its own workflow while using tools, and the tool calls kept working without misses. Setup: • Qwen3.8-Flash-Next • M5 Max • 128GB unified memory • MLX-Serve PR #363 • OpenCode 2 • 364.7K context The interesting part isn’t simply getting hundreds of thousands of tokens into memory. It’s what happens once the context gets this large. Long-context inference usually comes with a painful tradeoff. As the KV cache grows, memory pressure increases and generation can slow down. But this setup is still pushing through 364K tokens while maintaining a usable agent workflow. The model can reason, call tools, inspect results, continue working, and keep the session moving. And the tool calls reportedly haven’t missed so far. That’s important for agentic coding. A huge context window is only useful if the model can actually operate reliably inside it. A 400K-token context that constantly breaks tool calls isn’t very useful. A 364K session that can keep reasoning and executing tools is a different story. And the test isn’t finished yet. The current run is approaching 400K tokens, with the expectation that it can keep going. This is also another interesting example of why Apple Silicon keeps showing up in local LLM experiments. The M5 Max’s unified memory gives a large model and its growing KV cache access to one shared memory pool. With MLX-Serve continuing to improve, these machines are becoming surprisingly capable long-context inference boxes. The bigger takeaway: Context length is becoming a workload, not just a model specification. Running a model at 256K is one thing. Keeping an agent alive at 300K+ while it reasons and uses tools is much more interesting. And Qwen3.8-Flash-Next is showing that this can be pushed surprisingly far on a single 128GB Mac. 364.7K and counting. Next stop: 400K.

FHILY👑

39,982 просмотров • 9 дней назад

I asked Garry Tan how to use meta prompting to get better at AI: "My partners at YC Jared Friedman and Pete Koomen showed me how to do this. You can take almost anything that you do all the time and just drop it into a context window. And then say, “Here’s a bunch of inputs and outputs." And maybe you also add a bunch of notes. And then you tell it, “Write me a prompt that can act as an agent that takes this input and makes this output over here.” You can do this for almost any type of knowledge work. And you can even introspect. "What are things you notice that I did to convert this from the input to the output?”. And then you can just start using the prompt. Initially, it’s going to suck. Because it’s just not that smart yet. But what’s funny is now, I also use it to Iterate my writing. You can be very direct, "I would never say that", "Don’t say it like this", or "Oh, you used the long word there, use the short word". Just speak to it conversationally. And then when you're happy with the output, you can use that new output to make a new prompt. "Based on this conversation, give me a better initial prompt that incorporates all the things we talked about." And you can do this with literally everything. And in theory, there’s so much it applies to that people do day-to-day. You could use it for tweets. You could use it for editing podcasts. You can use it for pretty much everything. I have a folder of prompts that I use all the time. My YouTube prompt is on v27 or something. I'll go through this process with all the different max models. I'll use GPT 5.2 Pro. I’ll use Grok. I'll use Claude. Then, I’ll take all the outputs from all the models and put them into Claude and say "Here’s my prompt, here’s the output from four LLMs, including yourself. Rate each response and tell me what the pros and cons of each approach are." And I usually say "give it to me in numbered form". And then you can agree with one, disagree with two, tell it three is this or that. And then after that, you say given all of this, synthesize it."

I asked Garry Tan how to use meta prompting to get better at AI: "My partners at YC Jared Friedman and Pete Koomen showed me how to do this. You can take almost anything that you do all the time and just drop it into a context window. And then say, “Here’s a bunch of inputs and outputs." And maybe you also add a bunch of notes. And then you tell it, “Write me a prompt that can act as an agent that takes this input and makes this output over here.” You can do this for almost any type of knowledge work. And you can even introspect. "What are things you notice that I did to convert this from the input to the output?”. And then you can just start using the prompt. Initially, it’s going to suck. Because it’s just not that smart yet. But what’s funny is now, I also use it to Iterate my writing. You can be very direct, "I would never say that", "Don’t say it like this", or "Oh, you used the long word there, use the short word". Just speak to it conversationally. And then when you're happy with the output, you can use that new output to make a new prompt. "Based on this conversation, give me a better initial prompt that incorporates all the things we talked about." And you can do this with literally everything. And in theory, there’s so much it applies to that people do day-to-day. You could use it for tweets. You could use it for editing podcasts. You can use it for pretty much everything. I have a folder of prompts that I use all the time. My YouTube prompt is on v27 or something. I'll go through this process with all the different max models. I'll use GPT 5.2 Pro. I’ll use Grok. I'll use Claude. Then, I’ll take all the outputs from all the models and put them into Claude and say "Here’s my prompt, here’s the output from four LLMs, including yourself. Rate each response and tell me what the pros and cons of each approach are." And I usually say "give it to me in numbered form". And then you can agree with one, disagree with two, tell it three is this or that. And then after that, you say given all of this, synthesize it."

The Peel

51,632 просмотров • 6 месяцев назад

New short course: LLMs as Operating Systems: Agent Memory, created with Letta, and taught by its founders Charles Packer and Sarah Wooders. An LLM's input context window has limited space. Using a longer input context also costs more and results in slower processing. So, managing what's stored in this context window is important. In the innovative paper MemGPT: Towards LLMs as Operating Systems, its authors (which include the instructors) proposed using an LLM agent to manage this context window. Their system uses a large persistent memory that stores everything that could be included in the input context, and an agent decides what is actually included. Take the example of building a chatbot that needs to remember what's been said earlier in a conversation (perhaps over many days of interaction with a user). As the conversation's length grows, the memory management agent will move information from the input context to a persistent searchable database; summarize information to keep relevant facts in the input context; and restore relevant conversation elements from further back in time. This allows a chatbot to keep what's currently most relevant in its input context memory to generate the next response. When I read the original MemGPT paper, I thought it was an innovative technique for handling memory for LLMs. The open-source Letta framework, which we'll use in this course, makes MemGPT easy to implement. It adds memory to your LLM agents and gives them transparent long-term memory. In detail, you’ll learn: - How to build an agent that can edit its own limited input context memory, using tools and multi-step reasoning - What is a memory hierarchy (an idea from computer operating systems, which use a cache to speed up memory access), and how these ideas apply to managing the LLM input context (where the input context window is a "cache" storing the most relevant information; and an agent decides what to move in and out of this to/from a larger persistent storage system) - How to implement multi-agent collaboration by letting different agents share blocks of memory This course will give you a sophisticated understanding of memory management for LLMs, which is important for chatbots having long conversations, and for complex agentic workflows. Please sign up here!

New short course: LLMs as Operating Systems: Agent Memory, created with Letta, and taught by its founders Charles Packer and Sarah Wooders. An LLM's input context window has limited space. Using a longer input context also costs more and results in slower processing. So, managing what's stored in this context window is important. In the innovative paper MemGPT: Towards LLMs as Operating Systems, its authors (which include the instructors) proposed using an LLM agent to manage this context window. Their system uses a large persistent memory that stores everything that could be included in the input context, and an agent decides what is actually included. Take the example of building a chatbot that needs to remember what's been said earlier in a conversation (perhaps over many days of interaction with a user). As the conversation's length grows, the memory management agent will move information from the input context to a persistent searchable database; summarize information to keep relevant facts in the input context; and restore relevant conversation elements from further back in time. This allows a chatbot to keep what's currently most relevant in its input context memory to generate the next response. When I read the original MemGPT paper, I thought it was an innovative technique for handling memory for LLMs. The open-source Letta framework, which we'll use in this course, makes MemGPT easy to implement. It adds memory to your LLM agents and gives them transparent long-term memory. In detail, you’ll learn: - How to build an agent that can edit its own limited input context memory, using tools and multi-step reasoning - What is a memory hierarchy (an idea from computer operating systems, which use a cache to speed up memory access), and how these ideas apply to managing the LLM input context (where the input context window is a "cache" storing the most relevant information; and an agent decides what to move in and out of this to/from a larger persistent storage system) - How to implement multi-agent collaboration by letting different agents share blocks of memory This course will give you a sophisticated understanding of memory management for LLMs, which is important for chatbots having long conversations, and for complex agentic workflows. Please sign up here!

Andrew Ng

201,127 просмотров • 1 год назад

Right now, I have a few dozen tools to automate my life. Every single thing that I do more than once every week is now a tool: either a Claude Code skill, a scheduled workflow, or an application. • I spend the time doing this once. • I schedule it. • I forget about it (or at least, try to) CREAO is one of the platforms I've used extensively for building some of these automations. They are partnering with me on this post. They are one of the only platforms where you can go from a conversation to a scheduled agent that quickly. This is how it works: 1. You describe the problem you want to solve 2. CREAO's agent builds the logic and executes it 3. You can iterate with the agent to improve the solution 4. When you are done, you turn that solution into a mini app 5. You can schedule that app to run any time you need it Something important: These agents are deterministic. They don't use an LLM to generate answers, so they will always return the same output given the same input. This is critical: when you're automating something that runs every Monday at 9 am, you need it to return the same result every time, not a "creative" answer.

Right now, I have a few dozen tools to automate my life. Every single thing that I do more than once every week is now a tool: either a Claude Code skill, a scheduled workflow, or an application. • I spend the time doing this once. • I schedule it. • I forget about it (or at least, try to) CREAO is one of the platforms I've used extensively for building some of these automations. They are partnering with me on this post. They are one of the only platforms where you can go from a conversation to a scheduled agent that quickly. This is how it works: 1. You describe the problem you want to solve 2. CREAO's agent builds the logic and executes it 3. You can iterate with the agent to improve the solution 4. When you are done, you turn that solution into a mini app 5. You can schedule that app to run any time you need it Something important: These agents are deterministic. They don't use an LLM to generate answers, so they will always return the same output given the same input. This is critical: when you're automating something that runs every Monday at 9 am, you need it to return the same result every time, not a "creative" answer.

Santiago

13,727 просмотров • 5 месяцев назад

To me, microduck looks like a droid. So I made it sound like one Microduck is open source and hackable. That means you can program it to do whatever you want! I trained a 1.3m parameter LLM to output droid like responses. It is running in the browser The model has learned to respond with an emotion, alongside English like Happy, Excited, Sad Then the emotion is fed into a Rust based synthesizer to generate R2D2 like sounds that are aligned with the emotion --- completely procedurally, no wav files, no copyright violation You can chat with it. You can also customize it, give it a name and set your own colors! All set in a spaceship hangar backdrop with other friendly ducks---simulated with realistic physics and microduck's actual walking and skating policies! All in the browser, with WASM!!! I can put literally the SAME code in the browser, and then compile it into an ARM binary and put it in the actual robot in real life 🤯 If you like this and want to see more of it, please like and share so that my boss puts me in the priority list for the first microduck batch 😎 All open source of course! Microquack! 🦆 Hugging Face Space:

To me, microduck looks like a droid. So I made it sound like one Microduck is open source and hackable. That means you can program it to do whatever you want! I trained a 1.3m parameter LLM to output droid like responses. It is running in the browser The model has learned to respond with an emotion, alongside English like Happy, Excited, Sad Then the emotion is fed into a Rust based synthesizer to generate R2D2 like sounds that are aligned with the emotion --- completely procedurally, no wav files, no copyright violation You can chat with it. You can also customize it, give it a name and set your own colors! All set in a spaceship hangar backdrop with other friendly ducks---simulated with realistic physics and microduck's actual walking and skating policies! All in the browser, with WASM!!! I can put literally the SAME code in the browser, and then compile it into an ARM binary and put it in the actual robot in real life 🤯 If you like this and want to see more of it, please like and share so that my boss puts me in the priority list for the first microduck batch 😎 All open source of course! Microquack! 🦆 Hugging Face Space:

Onur Solmaz

33,500 просмотров • 17 дней назад

My RLM finally went recursive! Looking at these logs is way too addictive please send help. Notes: > Sent it 10 long wikipedia articles about deep learning (~2M context). > Asked it to find BLEU scores from Attention paper & explain MHA from these articles > RLM controlled by the new Minimax 2.5 ! Minor prompt changes were needed from the RLM paper. > Spends first 3 iterations understanding data format, works through errors, until it locates the Attention article from the mess. Like a human would use a Jupyter Notebook. > Launches subagent on only AIAYN article > This subagent launches 2 more subagents to fetch (a) BLEU score and (b) MHA (my original two-part question) > The lowest subagent returns the output using "FINAL_VAR" (i.e. it does not generate the text! Just finds the correct location in the context and sends it back as a variable) > Recursion propagates upwards > Outermost LLM recieves the RLM output, and generates the full text response. > Took 2.5 minutes walltime. Max recursion depth level was 2. 12 LLM calls in total. (This video contains cuts when the LLM is thinking/generating) > Subagents never gets to see more than 2000 characters. Only the outermost LLM gets to see the full output - it's needed to answer the final question, but its only 200-300 tokens compared to 2M! > Fully async. Code execution and subagent tasks can happen simultaneously! I feel soooo satisfied. Been some time since I've been this excited about shooting a tutorial video.

My RLM finally went recursive! Looking at these logs is way too addictive please send help. Notes: > Sent it 10 long wikipedia articles about deep learning (~2M context). > Asked it to find BLEU scores from Attention paper & explain MHA from these articles > RLM controlled by the new Minimax 2.5 ! Minor prompt changes were needed from the RLM paper. > Spends first 3 iterations understanding data format, works through errors, until it locates the Attention article from the mess. Like a human would use a Jupyter Notebook. > Launches subagent on only AIAYN article > This subagent launches 2 more subagents to fetch (a) BLEU score and (b) MHA (my original two-part question) > The lowest subagent returns the output using "FINAL_VAR" (i.e. it does not generate the text! Just finds the correct location in the context and sends it back as a variable) > Recursion propagates upwards > Outermost LLM recieves the RLM output, and generates the full text response. > Took 2.5 minutes walltime. Max recursion depth level was 2. 12 LLM calls in total. (This video contains cuts when the LLM is thinking/generating) > Subagents never gets to see more than 2000 characters. Only the outermost LLM gets to see the full output - it's needed to answer the final question, but its only 200-300 tokens compared to 2M! > Fully async. Code execution and subagent tasks can happen simultaneously! I feel soooo satisfied. Been some time since I've been this excited about shooting a tutorial video.

AVB

38,241 просмотров • 7 месяцев назад

someone just rebuilt claude code in pure python and made it 230x cheaper to run it's called clawcodex. a full from-scratch port of the claude code agent loop. 230k+ lines of python, MIT licensed it keeps your request prefix byte-stable, so DeepSeek's prompt cache covers your entire system + tools + history on every turn cache-hit input then bills at ~$0.0435 per 1M tokens the same 1M tokens on Claude Fable 5 runs you $10 so the longer you code, the more the cache pays off what you actually get: → 25 LLM providers (Claude, GPT, DeepSeek, GLM, local ollama) → 1M token context window → real agent loop: tools, skills, REPL, session history → one-line install, mac/linux/wsl and it's sitting at 860 stars. basically nobody is looking yet bookmark this before it gone

someone just rebuilt claude code in pure python and made it 230x cheaper to run it's called clawcodex. a full from-scratch port of the claude code agent loop. 230k+ lines of python, MIT licensed it keeps your request prefix byte-stable, so DeepSeek's prompt cache covers your entire system + tools + history on every turn cache-hit input then bills at ~$0.0435 per 1M tokens the same 1M tokens on Claude Fable 5 runs you $10 so the longer you code, the more the cache pays off what you actually get: → 25 LLM providers (Claude, GPT, DeepSeek, GLM, local ollama) → 1M token context window → real agent loop: tools, skills, REPL, session history → one-line install, mac/linux/wsl and it's sitting at 860 stars. basically nobody is looking yet bookmark this before it gone

alex

14,212 просмотров • 21 дней назад

someone just rebuilt claude code in pure python and made it 230x cheaper to run it's called clawcodex. a full from-scratch port of the claude code agent loop. 230k+ lines of python, MIT licensed it keeps your request prefix byte-stable, so DeepSeek's prompt cache covers your entire system + tools + history on every turn cache-hit input then bills at ~$0.0435 per 1M tokens the same 1M tokens on Claude Fable 5 runs you $10 so the longer you code, the more the cache pays off what you actually get: → 25 LLM providers (Claude, GPT, DeepSeek, GLM, local ollama) → 1M token context window → real agent loop: tools, skills, REPL, session history → one-line install, mac/linux/wsl and it's sitting at 677 stars. basically nobody is looking yet bookmark this before it gone

someone just rebuilt claude code in pure python and made it 230x cheaper to run it's called clawcodex. a full from-scratch port of the claude code agent loop. 230k+ lines of python, MIT licensed it keeps your request prefix byte-stable, so DeepSeek's prompt cache covers your entire system + tools + history on every turn cache-hit input then bills at ~$0.0435 per 1M tokens the same 1M tokens on Claude Fable 5 runs you $10 so the longer you code, the more the cache pays off what you actually get: → 25 LLM providers (Claude, GPT, DeepSeek, GLM, local ollama) → 1M token context window → real agent loop: tools, skills, REPL, session history → one-line install, mac/linux/wsl and it's sitting at 677 stars. basically nobody is looking yet bookmark this before it gone

alex

158,896 просмотров • 1 месяц назад

someone just rebuilt claude code in pure python and made it ~230x cheaper to run it's called clawcodex. a full from-scratch port of the claude code agent loop. 230k+ lines of python, MIT licensed it keeps your request prefix byte-stable, so DeepSeek's prompt cache covers your entire system + tools + history on every turn cache-hit input then bills at ~$0.0435 per 1M tokens the same 1M tokens on Claude Fable 5 runs you $10 so the longer you code, the more the cache pays off what you actually get: → 25 LLM providers (Claude, GPT, DeepSeek, GLM, local ollama) → 1M token context window → real agent loop: tools, skills, REPL, session history → one-line install, mac/linux/wsl and it's sitting at 677 stars. basically nobody is looking yet bookmark this before it gone

someone just rebuilt claude code in pure python and made it ~230x cheaper to run it's called clawcodex. a full from-scratch port of the claude code agent loop. 230k+ lines of python, MIT licensed it keeps your request prefix byte-stable, so DeepSeek's prompt cache covers your entire system + tools + history on every turn cache-hit input then bills at ~$0.0435 per 1M tokens the same 1M tokens on Claude Fable 5 runs you $10 so the longer you code, the more the cache pays off what you actually get: → 25 LLM providers (Claude, GPT, DeepSeek, GLM, local ollama) → 1M token context window → real agent loop: tools, skills, REPL, session history → one-line install, mac/linux/wsl and it's sitting at 677 stars. basically nobody is looking yet bookmark this before it gone

Argona

51,369 просмотров • 2 месяцев назад

Karpathy said something you'll regret ignoring: "We have to keep the AI on the leash. I'm still the bottleneck. I have to make sure this thing isn't introducing bugs and that there's no security issues." He said it at YC talk last year, when the worry was reliability. The models hallucinated and made mistakes no human would, so the leash implied keeping yourself in the loop and checking the output before trusting it. The models are far better now, and the line still holds, for a reason he was not focused on back then. Even a model that writes flawless code today still has no idea who is allowed to run it. Correctness and authorization are different problems, and only correctness improves as the model improves. A perfect agent still hands a tool where anyone can do anything, because permission was never part of the task. I actually tested this in practice with Claude Code. I asked it to build a small internal tool with a button that issues account credits. It worked first try, and running it locally, the credit applied the instant I clicked. Nothing decided who was allowed to click it. The agent wrote the right logic and displayed a success notification. It never checked whether the caller had the right, whether it should pause for a human, or whether anything was logged. And this is not a bug a smarter model can outgrow because the leash was never in the code. Identity, permissions, and audit live in the system that runs the app, not in what the agent generates. To solve this, I took the exact same bundle and hosted it on Retool. The credit write that fired silently on my laptop now stopped at an approval gate, resolved to a real identity through SSO, and landed in an audit log. I wrote none of it. The app inherited the entire boundary the moment it was deployed, and the video shows the before and after. You can try it yourself here: I also wrote a detailed breakdown of the whole thing in my recent article, and I worked with the team to put this together. It walks through the build, the exact moment the credit write went through on my laptop with nobody checking, and then what changed when the same app ran on Retool. It also covers why this is a property of the runtime and not something a better model fixes, which is why devs typically miss this. The article is quoted below.

Karpathy said something you'll regret ignoring: "We have to keep the AI on the leash. I'm still the bottleneck. I have to make sure this thing isn't introducing bugs and that there's no security issues." He said it at YC talk last year, when the worry was reliability. The models hallucinated and made mistakes no human would, so the leash implied keeping yourself in the loop and checking the output before trusting it. The models are far better now, and the line still holds, for a reason he was not focused on back then. Even a model that writes flawless code today still has no idea who is allowed to run it. Correctness and authorization are different problems, and only correctness improves as the model improves. A perfect agent still hands a tool where anyone can do anything, because permission was never part of the task. I actually tested this in practice with Claude Code. I asked it to build a small internal tool with a button that issues account credits. It worked first try, and running it locally, the credit applied the instant I clicked. Nothing decided who was allowed to click it. The agent wrote the right logic and displayed a success notification. It never checked whether the caller had the right, whether it should pause for a human, or whether anything was logged. And this is not a bug a smarter model can outgrow because the leash was never in the code. Identity, permissions, and audit live in the system that runs the app, not in what the agent generates. To solve this, I took the exact same bundle and hosted it on Retool. The credit write that fired silently on my laptop now stopped at an approval gate, resolved to a real identity through SSO, and landed in an audit log. I wrote none of it. The app inherited the entire boundary the moment it was deployed, and the video shows the before and after. You can try it yourself here: I also wrote a detailed breakdown of the whole thing in my recent article, and I worked with the team to put this together. It walks through the build, the exact moment the credit write went through on my laptop with nobody checking, and then what changed when the same app ran on Retool. It also covers why this is a property of the runtime and not something a better model fixes, which is why devs typically miss this. The article is quoted below.

Akshay 🚀

42,911 просмотров • 2 месяцев назад

🚨 YOUR GPU IS PROBABLY WASTING MORE THAN YOU THINK. vLLM is built to squeeze far more useful work out of your GPU when serving LLMs. Running an LLM at scale isn’t just about having a powerful GPU. The real problem is how efficiently you use its memory and compute. That’s where vLLM comes in. → High-throughput LLM inference and serving → PagedAttention for smarter KV-cache memory management → Continuous batching to keep GPUs busy → Prefix caching + chunked prefill → OpenAI-compatible API out of the box → Supports a huge range of modern LLM architectures → Quantization support for running models more efficiently And the crazy part? You can start an OpenAI-compatible inference server with: `vllm serve ` So your application can talk to your own model almost like it’s talking to OpenAI. The bigger idea: Don’t just buy more GPUs. Make the GPUs you already have work harder. That’s why vLLM has become such a major project in LLM inference. 🔥 #vLLM #AI #LLM #Inference #GPU #MachineLearning #AIInfrastructure #OpenSource #AIAgents #Developers

🚨 YOUR GPU IS PROBABLY WASTING MORE THAN YOU THINK. vLLM is built to squeeze far more useful work out of your GPU when serving LLMs. Running an LLM at scale isn’t just about having a powerful GPU. The real problem is how efficiently you use its memory and compute. That’s where vLLM comes in. → High-throughput LLM inference and serving → PagedAttention for smarter KV-cache memory management → Continuous batching to keep GPUs busy → Prefix caching + chunked prefill → OpenAI-compatible API out of the box → Supports a huge range of modern LLM architectures → Quantization support for running models more efficiently And the crazy part? You can start an OpenAI-compatible inference server with: `vllm serve ` So your application can talk to your own model almost like it’s talking to OpenAI. The bigger idea: Don’t just buy more GPUs. Make the GPUs you already have work harder. That’s why vLLM has become such a major project in LLM inference. 🔥 #vLLM #AI #LLM #Inference #GPU #MachineLearning #AIInfrastructure #OpenSource #AIAgents #Developers

Vikas gupta

14,138 просмотров • 4 дней назад

I cant believe this guy just made a permanent solution to context bloat and open sourced it all! when we tested this tool (Context+) for solving an issue on the OpenCode repository, the agent using this tool used ~6.5k fewer tokens, found the code and fixed it in half the time! the results were surprising: 6 to 10k tokens saved per prompt, completed task in ~2 minutes while the agent running without the tool took ~4 mins for the same and got stuck in loops bro built an entire beast by using all the modern tools that we could think of: undo trees, semantic search by meaning (by haskellforall), advanced refactoring, blast radius, advanced file context trees, restore points... i can keep going on semantic code search and context trees are the future of agentic coding and this tool proves it the feature i loved the most is semantic search and how it gets things done 2x faster with least possible tokens it makes an agent that actually knows what it’s doing and not just guessing, it makes meaning from your code similar to RAG. if you aren't optimizing your context, you are just burning money the developer says this tool is still under development, it can have unexpected behavior and the docs need updates but the video shows the reality of how fast it can be github: get here:

I cant believe this guy just made a permanent solution to context bloat and open sourced it all! when we tested this tool (Context+) for solving an issue on the OpenCode repository, the agent using this tool used ~6.5k fewer tokens, found the code and fixed it in half the time! the results were surprising: 6 to 10k tokens saved per prompt, completed task in ~2 minutes while the agent running without the tool took ~4 mins for the same and got stuck in loops bro built an entire beast by using all the modern tools that we could think of: undo trees, semantic search by meaning (by haskellforall), advanced refactoring, blast radius, advanced file context trees, restore points... i can keep going on semantic code search and context trees are the future of agentic coding and this tool proves it the feature i loved the most is semantic search and how it gets things done 2x faster with least possible tokens it makes an agent that actually knows what it’s doing and not just guessing, it makes meaning from your code similar to RAG. if you aren't optimizing your context, you are just burning money the developer says this tool is still under development, it can have unexpected behavior and the docs need updates but the video shows the reality of how fast it can be github: get here:

forloop

226,271 просмотров • 6 месяцев назад