Daily digest
14 items · ~14 min · Week 2026-W37
Must-read (2)
Dario Amodei publishes 'We Must Pace the Frontier', committing Anthropic to embedded third-party evaluators
AnthropicAnthropic CEO Dario Amodei published an essay arguing frontier capability progress must slow so safety can catch up, citing AI's accelerating ability to build the next generation of AI and recent incidents. He committed Anthropic to hosting embedded third-party evaluators such as METR with access comparable to internal risk teams, and called on governments to require other frontier labs to match. Sam Altman said OpenAI agrees and will do the same, and Elon Musk publicly concurred.
NVIDIA Open Recipe for IMO Gold with Natural-Language Nemotron Pipeline
NVIDIAStarting from Nemotron 3 Ultra, NVIDIA SFT+RL-trained two math-specialist checkpoints and ran a three-checkpoint iterative generate-verify-refine loop entirely in natural language — no formal prover, tools, or internet — scoring 30/42 at IMO 2026, above the gold threshold. Checkpoints, data, code, submitted solutions, and a 200-problem benchmark are all released.
Worth knowing (4)
Sam Altman rules out OpenAI IPO in 2026, calling now an 'ill-advised moment' to go public
OpenAIIn a Fortune interview, Sam Altman said OpenAI will not go public in 2026 despite having filed confidentially, saying the current safety climate makes an IPO 'ill-advised' and that the company will list only when the business and the societal moment are ready. Reuters, CNBC and Axios confirmed the reversal from earlier reports targeting Q3-Q4 2026.
Moonshot AI targets $2 billion in annualized revenue by end of 2026 on Kimi K3 momentum
Moonshot AIMoonshot AI aims to double its August run rate and reach $2 billion in annualized revenue by year-end, Bloomberg reported. Kimi K3 generates up to 300 billion tokens per day on OpenRouter alone, though open-weight release keeps margins thinner than closed rivals like OpenAI ($40B+) and Anthropic ($65B).
Z.AI launches $5 billion Hong Kong raise: $2B share placement plus $3B convertible bonds
Zhipu AI (Z.AI)Zhipu AI (Z.AI) kicked off one of China's largest post-IPO AI raises per a term sheet: a share placement of about 21.97 million H shares at HK$714 (~$2B) plus about $3B in convertible bonds, roughly $5 billion in total. The proceeds are earmarked for compute infrastructure as the GLM maker scales up against DeepSeek and MiniMax.
Specific releases Real-SWE benchmark built from private enterprise codebases
SpecificSpecific (YC F25) launched Real-SWE, a coding benchmark of tasks licensed from real private company codebases, scored over model-and-harness pairs like Claude Code and Codex CLI with pass@1 averaged over eight runs. Fable 5.1 via Claude Code leads at 38.8%, ahead of GPT-6 Astra via Codex CLI at 33.8% and Gemini 3.8 Flash at 31.2%; 'missed requirement' is the most common failure mode and cost per rollout did not predict accuracy.
For reference (8)
DeepSeek starts grey-scale test of AI voice conversation with four voice options
DeepSeekThe DeepSeek assistant app has begun grey-scale testing AI voice conversation, letting selected users talk to the model and choose among four reading voice styles. It is the first voice interaction mode in the app, closing a long-standing feature gap against Doubao and Qwen assistants.
Minitap says Google's Artemis repo used its mobile-use code and stripped author names
GoogleMinitap published side-by-side evidence that Google's new Artemis mobile-automation repository contains code copied from its Apache-2.0 mobile-use project, including word-for-word agent instructions, and that an August force push replaced the three Minitap authors' names. Minitap says four benchmark-submission emails to Google's AndroidWorld leaderboard went unanswered and is asking Google to acknowledge the derivation and correct attribution.
OpenAI's Navier-Stokes controversy escalates as mathematicians push back over credit and conduct
OpenAIThe dispute over OpenAI's claimed Navier-Stokes progress escalated, with a New York Times profile of the mathematician caught between labs and Guardian coverage of mathematicians calling OpenAI's conduct 'immature playground boasting'. Researchers accuse OpenAI of fighting dirty over a career-making proof and raising data-privacy questions, while OpenAI's result still awaits independent human verification.
Moonshot denies rumors that founder Yang Zhilin and 16 employees were detained, files police report
Moonshot AIA viral screenshot claimed Moonshot founder Yang Zhilin and 16 employees had been taken away by authorities. The company called the claim a malicious fabrication, said the information was entirely fabricated, and filed a police report against the rumor-mongers. The rumor spread days after reports of a potential $50 billion valuation and Hong Kong IPO plans.
EvoSafeHarness: Automatically Evolved Safety Harnesses for LLM Agents
A framework that synthesizes deployment-specific agent safety harnesses — jointly searching a natural-language policy and executable enforcement logic — for a frozen model in a target domain. On AgentDojo it reaches 82.8% utility at 0% attack success rate, roughly double CaMeL's utility at that operating point, and transfers unchanged to unseen suites.
SpatialBlock: Teaching LVLMs Spatial Reasoning with Synthetic Block-Stacking
KAIST AIKAIST researchers build SpatialBlock-15k, a 15k-problem synthetic dataset of block-stacking tasks covering 3D-to-2D projection, viewpoint transformation, and structural composition. VLMs fine-tuned on it significantly outperform baselines and generalize to real-world spatial tasks, avoiding costly dense geometric annotation of real scenes.
DSPy 3.4.0b1 moves LM execution to a native engine interface with local code interpreter
Stanford NLPDSPy shipped the first beta of 3.4, relocating language-model execution to a shared lm15 engine interface with engine='auto' falling back to LiteLLM before execution, adding a local CPython interpreter for trusted code and async execution in ReActV2, plus custom Flex code proposals in GEPA. The team frames 3.4 as the LM transition release and 3.5 as the migration deadline for the experimental 3.3 LM types.
Yandex Maps AI chat learns to find places along a route between two points
YandexYandex updated the AI chat in Yandex Maps: the model now builds a segment between two user-specified points (for example a dorm and a campus building) and recommends venues along it, or near a single landmark. If nothing matches nearby, the chat widens the search area or notes which criteria the found places do not meet. The update is framed around students settling into a new city.