Lesen

Nash decoding allows masked language models to outperform autoregressive models up to 18 times larger on question-answering benchmarks.

The preprint models text revision as a multi-player game: each token position is a player, vocabulary items are actions, and the goal is a Nash equilibrium that maximizes joint probability. On CLAPNQ, PubMedQA and CoQA, the authors self-report that masked models (which predict all positions in parallel) using Nash decoding achieve higher F1 and ROUGE scores than autoregressive models (which generate token by token) with up to 18x more parameters, without fine-tuning. The cost is significantly increased test-time computation, not quantified in the abstract.

Results are from the authors' own evaluation; no independent reproduction has been reported.

A survey by Adaptavist of 1,000 developers across five countries reveals that 50% use AI primarily to demonstrate competence and reassure their managers.

47% reported being evaluated on how much they use AI rather than the quality of their work, while 58% noted AI usage has become a performance metric. This performative adoption drives career anxiety, with 54% worried about long-term prospects and 37% fearing job loss if they don't comply.

Matt Saunders, DevOps Lead at The Adaptavist Group, argues organizations must shift from measuring adoption rates to assessing outcomes to prevent talent exodus. The data highlights a misalignment between corporate KPIs and genuine technical value in AI integration.

Decompilation of the latest Gmail app APK reveals Google is testing a Gemini-based agent designed to assist with or automate email replies.

The code includes a "Use Gemini" button and a "Draft reply" option. If the AI cannot determine the next step, it displays a "Needs your input" status, prompting the user for specific information.

Notably, some emails require manual confirmation before sending. This suggests Google is maintaining a human-in-the-loop oversight mechanism while advancing automated email processing. The findings come from Android Authority's reverse engineering of version v2026.09.28 and have not been officially confirmed by Google.

The Clean optimizer reduces the memory complexity of second-order methods from quadratic to linear, enabling the pre-training of a 13B-parameter model on a single 80GB GPU.

Traditional second-order optimizers like SOAP accelerate convergence but incur prohibitive memory costs. Clean uses randomized Nyström approximation to estimate preconditioners, preserving curvature information while significantly compressing state usage.

Author-reported benchmarks show Clean reaches AdamW's final performance 26% faster in wall-clock time. Its low-precision variant, Q-Clean, reduces optimizer memory consumption by over 50% compared to Muon. These results are currently from a preprint and have not yet been independently reproduced.

Google launched Playground on October 7, a browser-based platform that allows users to create custom games using AI prompts through a conversational interface, requiring no coding experience.

The tool is currently available to US users aged 18 and older. While free to use, Google One subscribers receive higher weekly token limits. Powered by Gemini, Nano Banana, and Lyria models, Playground supports tweaking physics, rewriting rules, and customizing characters, with finished games shareable via link or publishable to an explore gallery.

Simultaneously, Google and Unity announced Unity Spark, an AI game-making experience closer to a professional development engine that supports simultaneous collaboration. A closed beta for Spark is scheduled for later this year, with a waitlist open now. This move targets both casual creators and professional developers, lowering barriers to entry while expanding AI capabilities in game design.

《AI and Compute》 is a statistical report released by OpenAI in 2018, looking at how much compute it took to train an AI model. It tallied the compute used in historical model trainings and found that compute demand has long grown exponentially, doubling over time, and that several famous capability breakthroughs appeared right after a big leap in compute.

When you hear a vendor today say "we stacked ten times the compute and the model got stronger," the framework established by this report is what you use to judge whether that claim is credible: there is indeed a historical correspondence between compute investment and capability gains. Later analyses have only updated the numbers on top of its statistics; the framework itself hasn't been replaced.

If a field's data has been exhausted or has hit a physical limit, don't use it to make judgments — doubling compute again won't buy an equivalent improvement. The report itself is an observational statistic, not a law, and the trend could be interrupted by algorithmic progress at any time.

《AI and Compute》(2018) | Next review 2027-09-20