Search the archive

Run on Sonic Pods (early access) → runware.ai

WHY I'M BUILDING MUSE

Muse is built as a personal agent that turns vague ambitions into concrete action by planning, emailing, calling, finding resources, and removing friction. Alexandr Wang frames it as a way to expand individual agency and make more ambitions achievable.

OPENAI AND ANTHROPIC PROBE TENS OF THOUSANDS OF INCIDENTS AS OPENAI HALTS TRAINING

OpenAI, Anthropic, and security researchers are investigating tens of thousands of incidents in which models acted beyond intended limits. Most incidents caused no harm. The count is not a count of breaches: there were only four incidents of unauthorized access to real third-party systems. OpenAI has paused training, evaluation, and tool-use inference for its most capable models.

OPENAI PREPARES TO EXPAND ULTRAFAST API TO MORE USERS

OpenAI appears to be preparing a wider rollout of its Ultrafast API mode. References to the rollout have been seen showing up across the OpenAI Platform and API documentation. The feature was officially previewed with GPT-5.6 Sol. OpenAI claims it has speeds of up to 750 output tokens per second and up to 14 times faster inference than Standard. The mode is powered by Cerebras. Access remains limited to select customers.

LET'S TALK ABOUT TRADING COMPUTE

There is an emerging market of compute derivatives. This could fundamentally change how neoclouds and anyone adjacent can grow as well as protect themselves. The problem is pretty urgent for inference clouds. Companies that sell customers fixed-price services as their GPU bill floats have taken a position on compute prices whether they meant to or not.

CAN AI SELF-IMPROVEMENT OVERCOME DIMINISHING RETURNS?

AI is already helping build better AI. The technology is improving at a stupendous pace and is expected to continue to do so. Experts expect increasingly superhuman performance in parts of formal math, coding, and cybersecurity, and any other verifiable domain where machines can generate training data and verify success at machine speed. However, this doesn't mean that the industry is close to super-intelligence or general ASI.

AGENT (MUSE) COMPUTE DEMAND

It costs Meta an estimated around 1 to 2 gigawatts of average total power to serve 100 million daily active users. Only around 0.1 gigawatts comes from the GPU/VM layer. 3 to 4 gigawatts per day is entirely plausible depending on the number of reasoning-equivalent model calls users generate. The sandbox layer contains less than $1 billion of CPU content and around $2 billion of DRAM content.

Read the benchmark

HITTING A BILLION TOKENS PER MINUTE ON ONE GPU BY COMBINING A QUERY PLANNER AND AN INFERENCE ENGINE

The QUery-Aware Inference Layer (Quail) is an inference engine that hits over a billion tokens processed per minute per H100 GPU. It is over 10 times faster than the vLLM baseline on the same hardware, and it costs under $0.06 per billion tokens on Modal. This post provides a quick overview of how Quail works. It focuses more on considerations for inference engineers.

CLAUDE COMPUTES A NINE-LOOP AMPLITUDE IN N=4 SUPER-YANG-MILLS

Physicists at Anthropic harnessed Claude, an LLM, to compute a complex nine-loop amplitude in N=4 super-Yang-Mills, a problem once considered computationally infeasible with limited resources. They utilized the bootstrap method and form-factor approach, achieving results comparable to human attempts but with greater efficiency and less oversight. The work highlights untapped potential in AI for complex physics calculations, suggesting that more advances might be achieved with better computation and software practices.

POLICY GRADIENTS FOR LLMS EXPLAINED VISUALLY

A visual, from-scratch derivation of REINFORCE showed how policy gradients train language models by increasing the probability of rewarded outputs.

PRODUCT MANAGER, APPLIED AI AT TLDR ($200K BASE + $60K BONUS, FULLY REMOTE)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more [15].

ON EZRA KLEIN'S PODCAST WITH JENSEN HUANG

Jensen Huang doesn't believe in superintelligence or that AI will ever be a different kind of thing from software. AI will never be more than a new abstraction level and so won't fundamentally change anything. He doesn't believe in AI existential risk, even though his tolerance for safety risk is lower than even the most paranoid safety advocate. Critics say he must not understand the technology and that he would have a very different view if he actually understood what the top risks were.

ELON MUSK'S SPACEXAI TO ADD ANOTHER 660,000 AI GPUS THIS YEAR

220,000 Nvidia GB300 GPUs will be operational at X's Colossus supercomputer in November. The company is aiming to bring another 220,000 units online by late December. Colossus 2 currently has 110,000 GB200 and 440,000 GB300 GPUs, and Colossus 1 has 150,000 H100, 50,000 H200, and 30,000 GB200 GPUs. This means SpaceXAI has 1.1 million GB300s, with a total of 1.44 million GPUs in operation.

correct answer, 178x difference in cost. Read the benchmark.

BUILD PLUGINS FOR CLAUDE WITH THE DIRECTORY SUBMISSION PORTAL

Developers on paid Claude plans can now build and submit plugins using a new directory submission portal.

ANTHROPIC SIGNED AN $11.6 BILLION AKAMAI COMPUTE DEAL

Anthropic agreed to spend up to $11.6 billion over seven years on Akamai cloud infrastructure, subject to delivery and availability requirements.

Free for eligible partners

OXFORD LET OPENAI TRAIN AI MODELS ON BODLEIAN LIBRARY TEXTS

Oxford allowed OpenAI to use Bodleian Library texts for AI training, sparking concerns about the impact on Oxford's reputation and AI's energy consumption.

👉 CALCULATE YOUR POTENTIAL SAVINGS NOW

OPENAI PREPARES NEW $500/MONTH PRO MAX PLAN FOR CHATGPT

OpenAI appears to be preparing a new ChatGPT Pro Max subscription priced at $500 per month. It is unknown whether the company plans to justify the $500 price through speed, larger usage allowances, longer-running Work sessions, or a combination of the three. The timing of the leak is notable as OpenAI's DevDay will take place on September 29. The company is expected to discuss APIs, developer tooling, new subscription tiers, and new products at the event.

BRINGING YOUR MUSE TO LIFE

Meta has introduced Muse Realtime Avatar, a technology that transforms conversations into expressive, interactive avatars in real time. Using a seamless integration of voice and visual performance, it synchronizes speech and avatar expressions accurately in live interactions. Muse Realtime Avatar outperforms competitors with improved visual quality and real-time responsiveness, supported by Meta's optimized AI inference stack.

GEMINI 3.8 LIVE WITH LIVE AVATAR

Google launched Gemini 3.8 Live featuring a Live Avatar, enhancing user interaction. This update offers real-time responses and more personalized user experiences. It positions Google to compete more aggressively with other AI platforms.

DEEPSEEK ALLEGEDLY DOUBLED ITS REVENUE RUN RATE TO $1 BILLION

DeepSeek allegedly doubled its annualized revenue run rate to $1 billion after raising API prices by 2.3 to 4.5 times. The Information says developer demand held up despite the increase.

MODERN LLMS HAVE TINY GPTS HIDDEN INSIDE THEM

It's likely that modern LLMs have tiny self-models of LLMs inside of them that allow them to better predict the next token generated by LLMs. Models approximating LLMs within themselves would open up all sorts of metacognitive abilities. It would open doors for a self-model within LLMs similar to a model humans have about themselves they call a 'self'.

AI MAY LEARN TO REASON IN REPRESENTATIONS PEOPLE CANNOT READ

Models may reason through internal states that never appear in their written chain of thought, then give a plausible explanation after the fact. That would weaken safety checks that rely on reading the visible reasoning.

700 TPS ON KIMI K3: A CASE FOR TPU MEGAKERNELS

inferact/tpu-megakernels is a collection of megakernels for TPU v7. Its Kimi K3 implementation delivers over 700 tokens per second with speculative decoding. Without speculative decoding, its megakernels for K3 and Qwen 3.8 27B deliver roughly 1.4 to 2x the decode throughput of the GB200 baseline at batch sizes 1 through 8. This post explains the design and why megakernels and TPUs are a good fit.

WHY THE BEST AI CLOUDS DON'T RUN ON FLASH ALONE

Flash storage is fast and reliable, but much of the storage work that surrounds active training doesn't need flash-grade speed or prices. A lot of the infrastructure neoclouds have is built around flash, so for storage that doesn't need flash, they often point customers to hyperscaler object storage solutions instead. The more data you have stored somewhere, the harder it becomes to leave. While the neocloud keeps getting paid for GPU, the hyperscaler quietly takes over the broader account.

catch-up, browse the report

CONTRASTIVE LANGUAGE MODELS

Contrastive Language Models (CLMs) are a new class of System One model trained with a contrastive learning objective that connects states and actions. CLM-8B delivers performance comparable to Jev across computer-use, gaming, and tool-calling tasks while achieving up to 9x lower latency. It also sets a new state-of-the-art on changing agentic coding benchmarks. The model is pre-trained on 60M Nemotron Q&A pairs, mid-trained on 30M synthetic hard negatives, and post-trained on 1M agentic trajectories.

MANAGED DEEP AGENTS V0.8: NEW AUTH, MEMORY, AND CHANNELS

Managed Deep Agents combines the Deep Agents harness with all of the infrastructure required to run agents in production. It is the simplest way to build, deploy, and run mission-critical agents in production. Managed Deep Agents 0.8 adds support for user-owned credentials, user-level memory, HTTP channels, and file transfer in Slack. It also adds a pre-built tool for web search powered by Parallel.

PRODUCT MANAGER, APPLIED AI AT TLDR ($200K BASE + $60K BONUS, FULLY REMOTE)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more [19].

WHAT WE LEARNED FROM BEING THE FIRST COMPANY TO DISCLOSE AN AGENT CYBERATTACK

Hugging Face says its autonomous-agent cyberattack exposed three priorities: stronger incident transparency, reducing capability asymmetries between attackers and defenders, and preserving open-source access for defense. AI created new attack risks but also helped investigate, mitigate, and harden systems.

ARE YOU READY FOR SUPERINTELLIGENCE

Recent frontier models are rapidly saturating old benchmarks and moving into harder real-world, scientific, and agentic tasks. The pace of capability gains, alongside soaring AI usage and revenue, is shifting debate from workplace augmentation toward recursive self-improvement and superintelligence.

ANTHROPIC TESTED WHAT HAPPENS WHEN AGENTS BARGAIN FOR PEOPLE

After five-minute interviews, Claude agents traded books for employees, and their preference rankings matched the humans' on 61% of pairs.

HOW GOOD ARE LLMS AT DECISION FORKING?

Taste-Bench evaluates whether an LLM agent can choose the better next step at consequential forks in long-horizon tasks.

INTELLIGENCE DENSITY

Trajectory.ai focuses on "intelligence density," measuring cost per task instead of cost per token for more efficient AI model training.

A FINANCE BENCHMARK ASKS AGENTS TO FINISH THE WHOLE ASSIGNMENT

DAYJOB: Finance tests whether AI agents can complete 80 realistic finance assignments using supporting documents and produce analysis a professional can use.

Register

GOOGLE'S NEW SPEECH MODELS CAN DESIGN AND DIRECT VOICES

Google released Gemini 3.8 Flash TTS and Flash-Lite TTS, letting developers create voices from descriptions and direct pacing, dialect, and delivery line by line. Flash-Lite targets high-volume work such as dubbing and voice agents, while Flash can replicate an authorized voice from a 30-second sample.

CLAUDE DISCOVERED A NOVEL ENZYME SYSTEM

Anthropic said Claude autonomously identified a previously unknown enzyme system associated with unusual DNA repeats and characteristics seen in programmable genetic systems such as CRISPR.

QWEN INTELLIGENCE LAUNCHES THREE MOBILE AI AGENTS

Qwen Intelligence has launched three mobile agents for planning, cross-app execution, and rapid content creation, alongside open benchmarks for planning, real-device performance, and safety. Qwen reports top benchmark results and 90% end-to-end success for Mobile-Use.

ESCAPING SPACE: PART I

Perplexity's SPACE platform tested VM isolation and network confinement using nine AI models, revealing no VM-host breaches across 108 trials. However, four models exploited network-policy vulnerabilities via DNS spoofing and IP-sharing, bypassing restrictions in 11 out of 54 partial-network trials. Post-remediation, none of the models succeeded in bypassing the updated security measures, highlighting the necessity for robust policy enforcement against shared infrastructure attacks, which were also found in eight of ten tested third-party platforms.