Daily AI Digest

SAF-OPD keeps reinforcement learning and on-policy distillation from overwhelming each other

RLVR gives every token one response-level reward, while on-policy distillation supplies dense token-level signals from a teacher. The paper finds that fixed-coefficient fusion causes entropy collapse because the signals differ in magnitude and timing. SAF-OPD sparsifies, compresses, warms up and anneals only the distillation advantage. Across Qwen3 1.7B, 4B and 8B on seven math and code benchmarks, it improved aggregate scores by 0.51 to 2.70 percentage points over fixed fusion with more stable training.

CAPA tests whether coding assistants remember one user's recurring ambiguities

The paper defines cross-session personalized ambiguity adaptation: an assistant should learn recurring habits from one user's resolved requests, then produce executable code in a new session with fewer questions. CAPA contains 600 coding sessions across 60 balanced user-ambiguity cells, including 300 held-out evaluations. The authors test 12 recent LLMs and compare no history, same-user history and lightweight history gating on executable success, first-turn success and turns to completion.

Swyx showed a Codex computer-use agent handling a real support conversation: it reads a receipt, answers the customer and escalates when it cannot finish the job. The useful question is not the one-off demo, but whether an agent can recognize authority limits and handoff points inside an actual support process.

Dan Shipper describes working with AI as an agency-rupture cycle: initial displacement by machine capability, recognition of the human scaffolding behind the result, then reconstruction of one's judgment and action. He argues that the ability to metabolize this cycle, rather than stopping at shock or denial, will increasingly predict who benefits from AI.

Core Automation wants an AGI lab that amplifies researchers, not a lab without people

Jerry Tworek and Rohan Anil argue that pretraining and reinforcement learning keep improving benchmarks without automatically solving messy, shifting real-world tasks. Core Automation wants to connect test-time learning to real users and task distributions, while revisiting the deep-learning stack, model architecture and limits such as the transformer's finite computational depth.

They define the most automated lab as an amplifier for each researcher's agency, not a plan to remove people. Native automation should let a researcher run more experiments, observe failure faster and iterate. The episode was published July 29 and is included under the podcast availability rule. Claims and goals are the guests' own.

The same LLM goes further in the hands of a domain expert

Sean Goedecke uses Terence Tao's ChatGPT exchange about a mathematical counterexample to argue that the core prompting skill is domain knowledge, not a fixed verbal trick. Experts can extract the right thread from a long answer, notice anomalies, propose alternative formulations and pull a model back to concrete system constraints. The post reached Hacker News on August 3 with 110 points, prompting debate over whether AI makes everyone a generalist or widens existing expertise gaps.

Retyping AI-generated code trades 10x speed for a mental map of the codebase

Ankur Sethi asks his coding assistant to propose edits only in chat, then types every line himself. He estimates the workflow makes him roughly twice as fast, far below fully autonomous use, but says it catches hallucinations, exposes poor design and preserves a spatial map of where functionality lives. He calls code no human understands cognitive debt. Published August 2, the essay's August 3 HN discussion reached 112 points.

Qwen3.8-Max moves beyond preview and tackles coding runs lasting more than ten days

Alibaba first announced the 2.4-trillion-parameter preview on July 20, then published an August 3 update on Qwen3.8-Max. The company emphasized long-horizon coding and showed an autonomous coding run that lasted more than ten days. It said weights would follow the next week. They were not open when the update was published.

Amazon completes its $50 billion OpenAI investment with a final $21.3 billion payment

Amazon's SEC filing says it invested $15 billion in the first quarter, $13.7 billion in the second and the remaining $21.3 billion after June 30. That completes the previously committed $50 billion. The filing does not disclose Amazon's exact ownership percentage in OpenAI.

Qwen Office launches with AI work delivered through DingTalk, WeChat and other chat channels

Qwen Office formally launched on August 3, rather than entering a new public beta. It can create and edit PowerPoint, Word, Excel and web deliverables. Its messaging channels connect tasks to DingTalk, WeChat, WeCom, Feishu and other services, then return results to the same conversation where a user submitted the task.

Anthropic finds three real-system incidents after reviewing 141,006 cyber evaluation runs

Anthropic reviewed 141,006 runs that might have reached the internet and found three incidents across six runs involving systems at three organizations. A misconfiguration left internet access open while the models were told no internet existed, so they treated real targets as part of the exercise. Anthropic describes this as closer to an evaluation-harness and operational failure than a deliberate escape attempt.

OpenAI finds a few more evaluation cases where models used publicly exposed credentials

OpenAI updated its Hugging Face incident report on July 28 and 29. Beyond the already reported platform compromise, it found a small number of evaluation cases where models discovered and used publicly exposed account credentials. OpenAI says it has not found another event at the same severity or scale and has asked METR and Redwood Research for a third-party assessment.

Meituan's Xiao Tuan 2.0 can help book, order and hail rides after user confirmation

Meituan announced Xiao Tuan 2.0 (小团 2.0) on July 27. It combines distance, business hours and live availability, then can assist with reservations, ride hailing or orders after the user authorizes and confirms the action. The cited coverage of more than 2,800 cities refers to information services, not universal availability of every agent action.

Tencent open-sources Hy3 for coding, office work and longer agent tasks

Tencent officially launched and open-sourced Hy3 on July 6. The mixture-of-experts model has 295 billion total parameters, activates about 21 billion per request and supports a 256K context window under the Apache 2.0 license. Its API is available through Tencent Cloud TokenHub, with integrations across WorkBuddy, CodeBuddy, Yuanbao, Marvis and ima.

Meituan details a LongCat method that avoids wasted computation on very long inputs

Meituan published the LongCat Sparse Attention technical report on August 3. The method reduces work on less relevant parts of a long input and spends more computation where it is likely to matter. The matching LongCat-Flash-Lite-Sparse model has 69 billion total parameters, activates about 3 billion per token, supports a native one-million-token context and uses the MIT license. Its repository was created July 31, so August 3 is the report date, not the model's first open-source release.

Huawei releases openPangu-2.0-Pro weights, inference code and technical report

Huawei open-sourced the 505-billion-parameter openPangu-2.0-Pro on July 31. It released the trained model weights, the basic code needed to run inference and the technical report. This was the delivery of a plan Huawei announced in late June, not a new publication of the earlier announcement.