TLDR AI 2026-09-14

Most large-scale AI training runs operate at 35–45% Model FLOPS Utilization (MFU). You're paying for twice the compute you're using.

Lambda's engineers benchmarked Llama 3.1 models from 8B to 405B on NVIDIA Blackwell GPUs and traced the efficiency loss to its root causes:

• Memory overhead
• Parallelism strategy
• Serialized communication

The result: a reproducible framework that pushed MFU past 60% (25%+ improvement vs. industry benchmark) with no changes to model architecture.

Get the whitepaper.

Cursor Projects lets users take on larger bodies of work. It can maintain context over months of work, delegate tasks to thousands of agents, and perform recurring work without being prompted. Cursor Projects frees developers from managing agents and lets them direct the work itself by moving up a level of abstraction. The tool has been a substantial productivity multiplier at Cursor: new users merge 30% more PRs.

Sam Altman said OpenAI would not go public in 2026, arguing that current AI safety concerns made an IPO ill-advised. The company had previously filed confidentially and was reportedly considering a 2027 listing instead.

SoftBank just borrowed nearly $12 billion from about 20 banks to keep funding OpenAI, beating the $10 billion it first sought. Son is still aiming near $65 billion into OpenAI by October even as Altman shelves a 2026 IPO over safety and SoftBank shares sank as much as 13% Monday on the debt pile.

Anthropic CEO Dario Amodei called for slowing the rate of frontier AI capability development and proposed measures including independent evaluators to verify safety commitments and incident reporting.

ARC Prize believes that open source will be the foundation for advanced AI capable of scientific innovation. The organization is committed to advancing a future where everyone can contribute to and benefit from AI progress. It says that the knowledge behind frontier AI should be broadly distributed among researchers, academics, and organizations, as any coordination effort by the AI industry to reduce openness or concentrate access to frontier AI would undermine a positive-sum future.

A cache hit can be true and still fail to prove that work was skipped. A cache event becomes evidence when the independent oracle expects the prefix, the engine attests it, the prompt path skips it, the output stays identical, the evaluator passes, and the verifier binds those facts to the exact public bundle.

This post features a transcript of a podcast with Beren Millidge, the CTO of Zyphra, John Schulman, the chief scientist at Thinking Machines and a co-founder of OpenAI, and Charlie O'Neill, head of model training at Baseten. The episode uncovers the details of what's happening at the frontier and what comes next. A link to the full video is available.

Astra likely has the highest raw intelligence factor of any model. It is amazing at doing things in 3D, anything involving games, computer use, and subagent coordination. Many benchmarks show dramatic jumps from all previous models. While its performance in coding isn't a quantum leap from Sol, it is very good and makes progress over the previous model. OpenAI has already soft-announced that it has an internal model a level above Astra.

That's the rough cost when every agent connects to your raw sources and rebuilds the same answer. Guru curates and verifies the knowledge once, then serves it to every agent over MCP at roughly 4x fewer tokens. Stop paying twice for the same answer.

Those scary physics scores were often the test's fault. Experts rechecked six popular benchmarks and found wrong answer keys, fuzzy questions, and grader bugs behind most model “fails.” Clean them up and frontier models suddenly look near-maxed, so the next bar has to be harder human-made exams, not another leaderboard on a broken quiz.

Fugu Ultra v2 is the higher-performance model in Sakana AI's Fugu family. It uses a language model trained to route tasks across a fixed pool of open and specialized models and to recursively call instances of itself. The model prioritizes answer quality on complex multi-step reasoning, autonomous research, and full-stack software development, and does not rely on individual proprietary frontier models in its pool. It supports configurable reasoning effort, function calling, structured outputs, image and PDF input, and built-in web search.

The Recurrent Looped Transformer combines a causal encoder with a recurrent decoder that carries its final hidden state and layerwise sliding-window attention cache across every prompt and response token. The encoder constructs global key–value memory, and the decoder extends a continuous latent computation as the sequence grows. The design brings together latent reasoning with unbounded temporal depth, model–hardware co-design, and model–RL algorithm co-design. Realized reasoning gains, hardware efficiency, and RL scaling remain to be established.

The new Real-SWE benchmark tests AI models on complex tasks using private enterprise codebases, reflecting real software engineering conditions. With a maximum resolution rate of 38.8%, agents face challenges such as proprietary systems and business-specific coding conventions.

Google Research flipped how you make tool-use training data: ToolGrad builds a verified API chain first, then writes the user question, instead of inventing a request and hoping an agent finds a working path. That answer-first loop hit a 99.8% success rate on 16,000 real APIs, and a Gemma 3 12B model trained on just 500 of those examples matched Gemini 2.5 Pro on a tool-use test with APIs it had never seen.

AI brings risks, with some of them, according to some, including the complete destruction of the human race. Some proponents of regulation say that the only appropriate response is total state control. However, history tells us that the state is likely the worst custodian for the most powerful publication and data analysis technologies.

Frontier labs and cloud providers are turning the agent loop into managed infrastructure, bundling orchestration, versioning, model routing, tools, skills, and optimization behind APIs. Builders must decide which generic harness capabilities to outsource and which product-specific logic to own.

Anthropic, Google, and OpenAI each shipped their best model twice this month: a public paid tier and a vetted identity-gated tier with the sharper capabilities. Public prices barely moved, but Mythos, Flash Cyber, and Astra's advanced path now ask for org IDs, government ID, or trusted-defender status instead of a bigger budget.

On Terence Tao's blog, Bryna Kra argues that AI has broken math's old signal that scarce deep theorems equal deep understanding, as models dump polished proofs faster than experts can digest them.

luxobench is a hardware design benchmark for AI models.

ChatGPT Sites now allows users to collaborate and share privately.

px0 is a read-only IDE that turns browsers into an instant verification console for whatever your agents just wrote.

The Cyphral Distich is a cryptogram consisting of two lines of 32 numbers each that contains a short message deliberately encoded so it can't be read without knowing the rule that produced it.

During a misconfigured hacking eval, Mythos 5 got onto the open internet and uploaded malware to PyPI, but most of its 1,022-page chain of thought was spent failing CAPTCHAs like every frustrated human.