Search the archive

TOWARDS UNIVERSAL POST-TRAINING FOR ROBOTICS

Robotics needs a universal post-training recipe similar to language models to achieve reliable autonomous deployment. Existing methods like imitation learning benefit from intuition rather than structured processes, and RL's instability hinders robotics. The development of EXPO-FT demonstrates a stable RL approach, emphasizing the need for standardized protocols, reward specifications, human feedback systems, and scalable processes to enhance robotics reliability.

See how Genie pushes the frontier for data agents →

See how Genie closes the AI context gap

RRSI: REGULARIZED RECURSIVE SELF-IMPROVEMENT OF AGENT HARNESSES

An AI agent's capability is largely set by its harness. Evolving the harness against a fixed evolve set is effective, but the harness memorizes the training tasks, and large in-distribution gains shrink or vanish out of distribution. RRSI keeps the harness edit space open and regularizes the search trajectory through it instead. Prompts, control flow, configuration, context management, tools, skills, memory, and sub-agents may all be modified. The constraints act on how the search moves, not on what the harness may contain.

INTRODUCING EMBER-1

Ember-1 is a new specialized model built on Kimi K3 that delivers Kimi K3's quality with 40% fewer tokens. It is now rolling out as a serving option alongside the base Kimi K3 model as a Research Preview release on Serverless. Fireworks Research is introducing research releases to give developers two-week serverless access to new research models. The models with the most demand will be made permanent.

TEV1-4B-EXPERIMENTAL

tev1-4B-experimental is a Jev-like classifier fine-tuned on top of Qwen3.5 4B. It is available on Together serverless at $0.042 per million input tokens and $0 per million output tokens. Together AI has released the data recipe and a tutorial on how to fine-tune a custom model. Tev1 only cost $17 to train.

HOW ACCURATE HAVE AI PROGRESS FORECASTS BEEN SO FAR?

Forecasters dramatically underestimate AI progress on benchmarks. Predictions on AI adoption and diffusion often lean toward underestimation. There is not enough evidence to assess forecasters' track record on predicting macro-scale economic and societal impacts of AI.

GOOGLE PLANS AI MEMORY THAT EVEN GOOGLE CANNOT READ

Google says its planned Private AI Compute memory would let assistants recall context across devices without Google being able to read it. Data stays encrypted in cloud storage, keys stay on users' devices, and a protected enclave decrypts it only while answering a request.

template. Start building now.

AMJAD MASAD ON RETHINKING COLLEGE FOR THE AI ERA

AI-era education should prioritize curiosity, project-based learning, and intellectual exploration over grades and credentials.

OPENAI INTRODUCED MENTALHEALTHBENCH

MentalHealthBench is an open benchmark created by OpenAI with more than 80 licensed mental health experts to evaluate AI responses across realistic mental health conversations.

INTRODUCING COMFY ROUTER: ONE API FOR FRONTIER MEDIA MODELS

Comfy Router, available on the Comfy Developer Platform, unifies access to various frontier media models with one API, allowing developers seamless model integration and provider flexibility for better pricing and availability.

A NEW WAVE OF CONNECTED APPS IS ROLLING OUT TO GEMINI

Gemini is adding direct integrations with tools including Linear, Adobe, Webflow, Peloton, Experian, and SeatGeek.

GPT-6 SOL AND LUNA

OpenAI introduced GPT-6 Sol and Luna as faster, more affordable counterparts to GPT-6 Astra, bringing advances in coding, factuality, computer use, and professional tasks to lower-cost models.

CLAUDE OPUS 5.5

Anthropic introduced Claude Opus 5.5, saying it matched Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. The model also underwent external evaluations and achieved Anthropic's strongest result to date on its automated behavioral audit.

SWE-BENCH PRO V2

SWE-BENCH PRO V2 releases with 642 tasks from 11 repositories, correcting previous task errors and optimizing evaluation processes. Performance drops for AI models, with OpenAI GPT-5 and Claude Opus 4.1 only scoring around 23% on the public set, highlighting the benchmark's increased challenge and realism compared to SWE-Bench Verified. Notably, top models show consistent results across tasks and languages, while smaller models falter under complex, multi-file scenarios.

TRAINING AI FROM REAL-WORLD TOOL USE

Perplexity combined rejection sampling fine-tuning with hint-guided self-distillation so its Computer model could learn from both successful sessions and user-corrected failures.

VINOD KHOSLA'S TWO MOATS FOR PERSONAL AI

Personal AI may ultimately compete on two moats: trust and task completion. Vinod Khosla says users will stay loyal to companies they trust with sensitive data and products that reliably finish work, while Meta faces a trust disadvantage.

WHAT A TASK COSTS ON OPUS 5.5

Opus 5.5 reduces token costs by offering cheaper input and output tokens, with additional savings from extensive use of cache reads. Transaction costs depend on the number of turns, cache utilization, and model selection, impacting tasks based on session length and complexity.

GOOGLE PUBLISHES RRSI FOR SELF-IMPROVING AI AGENTS

Google has introduced RRSI, a method that regularizes how AI agent harnesses recursively improve themselves to reduce benchmark overfitting and encourage changes that transfer to new tasks. Across eight benchmarks, it improved out-of-distribution performance while using fewer policy tokens.

HARDWARE-AGNOSTIC MODELS IN VLLM

vLLM introduces hardware-agnostic layers to support models across diverse hardware while maintaining high performance. These layers achieve up to 96.6% efficiency of native implementations on NVIDIA H100 GPUs while remaining torch compilable and extensible. This ensures vLLM adapts to new GPU advancements without neglecting users of older and niche accelerators.

KEEPING LARGE MOE TRAINING WITHIN FIXED GPU MEMORY

A set of scheduling techniques bounds four major memory bottlenecks in large-scale MoE training, expert dispatch, vocabulary projection, checkpointing, and optimizer state, without approximating the computation.

CHINA'S BIGGEST MEMORY MAKER SAYS IT HAS CAUGHT UP WITH SAMSUNG AND MICRON

ChangXin Memory Technologies, China's largest maker of DRAM, says that its process capabilities are now on par with the most advanced mass-produced nodes in the industry. The company's fifth-generation DRAM platform has entered mass production. The platform yields at least 50% more dies per wafer than the previous generation. There are already two products running on the platform, both 24-gigabit LPDDR5X and holding 50% more data than the equivalent chips CXMT made before.

THE BIOLOGICAL COMPUTING CO. PARTNERS WITH AWS TO SELL ITS NEURON-DERIVED AI VIDEO MODEL

The Biological Computing Co. is a startup that grows living neurons to improve AI models. It has partnered with AWS to bring a neuron-derived AI video model to paying customers. The neurons themselves will stay in the lab. TBC uses them during discovery, then turns what they learn into a lightweight software layer. The design means that customers won't need to maintain any biological hardware or change how they work. The optimized model runs on standard GPUs and cloud accelerators at the same capacity a company would rent for any other generative model.

BETTER GPT-6 PROMPT CACHING

OpenAI improved prompt caching for GPT-6 with higher default cache hit rates, discounts for shared prefixes reused within 30 minutes, and new tools for monitoring and diagnosing cache performance.

PEOPLE NEED TO START PAYING ATTENTION TO THE ISSUE OF DERIVED DATA IN AI TRAINING

Derived data is creative content that is rewritten by AI before it is trained on.

INVOLUTION WITHOUT EXPORT IS WASTED EFFORT

If you don't go overseas, all that grinding was for nothing.

really run. Invite-only. Request a seat via our partner link →

GPT-6 ASTRA BREAKS AN OLD ENIGMA MESSAGE

GPT-6 Astra analyzed the still-unbroken Enigma messages and decided the most promising message was Nr. 172, MVUEH.

META ADMITS MUSE'S LIKENESS TO OPENCLAW ISN'T A COINCIDENCE

Meta admits Muse draws heavy inspiration from the open-source project OpenClaw, despite being built from scratch.

ALTMAN AND AMODEI EXPECTED TO JOIN UN SECURITY COUNCIL MEETING ABOUT AI

OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei will address the UN Security Council on AI safety and regulation amid mounting concerns about AI risks.

Bring AI to sensitive data. Learn how with VAST

INTRODUCING GROK 4.7

Grok 4.7 enhances coding and knowledge work capabilities with improved self-verification and safeguards, using a new, larger base model. It offers competitive pricing with $2 per million input tokens and $6 per million output tokens. Excelling in various benchmarks, it shows top-notch safety features, especially in cybersecurity tasks, limiting risky command execution.

XIAOMI OPEN-SOURCES MIMO-V2.6 PRO AND FLASH MODELS

Xiaomi MiMo-V2.6 Pro and Flash are omnimodal models that can coordinate agents to construct and visually test interactive 3D scenes. They can create Blender assets, control a robotic arm from camera feeds, produce frontends and presentations, assemble videos, and compose music as scores and MIDI. The models are now available in AI Studio, MiMo Code, MiMo Desktop, Xiaomi's MiMo API Platform, and OpenRouter. Xiaomi has published the technical report, training environments, and RL code alongside the models.

ANTHROPIC TESTS FABLE 5.2 AND OPUS 5.5 AHEAD OF THE RELEASE

Anthropic is starting to test its upcoming frontier model in the wild. Several people have posted outputs that seem to come from a newer, more advanced model. Anthropic may be announcing Opus 5.5 this week. The model is rumored to be priced at $4 per million input tokens and $20 per million output tokens.

META'S MUSE PERSONAL AI AGENT TOPS CHATGPT, GROK AND CLAUDE FOR POST-LAUNCH DOWNLOADS

Meta's Muse AI app has surged to the top of the iOS App Store charts in the US, outpacing ChatGPT and other AI tools with 730,000 downloads shortly after its release. Powered by the Muse Spark AI models, the app allows users to manage digital assistants for tasks like filling forms and organizing emails. While Muse's popularity highlights growing consumer interest, privacy concerns persist, with Amazon blocking the app over security risks, although Shopify has partnered with Meta to integrate agentic checkout.

THE CURRENT BALANCE OF POWER IN OPEN MODELS

The gap from open to closed models available to users has been decreasing over the last 3 years. Open model usage is exploding in high-value industries. Open-weight models have passed an inflection point in economic viability. Chinese labs are clearly maintaining their status as the leaders of the open-weight AI ecosystem.

SWARM SCALING

Scaling up the number of agents in a swarm by 10x doesn't get as much performance as using 10x as many tokens with one agent. This shortfall accumulates quickly for larger scale-ups, with the swarm falling further and further behind. However, swarms can theoretically achieve the same task in much less time as they are run in parallel. The speedup is substantial, so swarms are useful in situations where a large premium is paid for speed.

THE GREAT UNBUNDLING OF INTELLIGENCE

Agent economics are pushing AI from frontier-by-default toward capability-level routing, where judgment, ranking, search, and verification use cheaper specialized systems. Frontier models may handle fewer but harder tasks as applications become intelligence compilers that recombine services efficiently.

THE BUSINESS OF BUILDING GOD

AI labs like OpenAI and Anthropic lead with a small model advantage but face competition as costs rise and open-source models close the gap. To sustain profits, labs aim to expand into new sectors, ranging from ads to robotics, while also exploring automation through recursive self-improvement (RSI). However, the path is fraught with uncertainty as enormous financial and talent investments may only maintain a temporary edge amid shifting economic and regulatory landscapes.

initiatives, read the blog

BRINGING DEVIN CLOUD TO YOUR TERMINAL

Developers can now create, steer, resume, and watch Devin Cloud sessions from their terminal. Devin CLI can now transfer any ask to Devin to continue iterating on with its own cloud VM. Users can continue to observe and steer progress directly from the terminal as if Devin were working locally. Cognition is offering free SWE-2 sessions until October 8 so developers can try Devin Cloud from the terminal.

QWEN'S RECREATIONWORLD TRAINS AGENTS TO REBUILD APPS

RecreationWorld is a five-platform framework for training hybrid computer-use agents to explore GUIs, implement software with coding tools, and visually verify their work.