Alphabet loses about $475 billion in a day over AI spending; also: a record $6.3 billion Binance theft, A-RSI self-improvement, an agentic RAG test where the agent lost, and Co-Design Bench

Lead stories

Alphabet's worst day since its IPO, and the week the market questioned AI spending

1 source, 5 min read

What happened. On Thursday, October 29, Alphabet shares fell about 13% the day after its third-quarter results. Google Search revenue grew more slowly than expected, and free cash flow turned sharply negative. Morning Brew, citing the loss as about $475 billion of market value, summed up the week's earnings as "A lot of money is going into AI and not enough is coming out." Amazon fell 1.41% on Friday after reporting higher-than-expected AI spending and slower retail growth. Intel fell 28.58%, its worst day in 50 years, after weak PC demand hit its outlook. yahoo

Details. The size of the loss depends on who you read. Yahoo Finance puts it at $475 billion and has corrected an earlier version of its story. Its correction note says the earlier version "misstated Alphabet's one-day loss in market value on Oct. 29. It was $475 billion," although the article URL still says $500 billion (Yahoo Finance). Fortune says Class A shares fell 12.6% to $283.82 and wiped out about $482 billion, and that the previous worst day was an 11.6% drop in March 2020 (Fortune). CNBC rounds it to about $500 billion (CNBC). Yahoo ranks $475 billion as the second-largest one-day loss for any US-listed company, behind Nvidia's roughly $593 billion on January 27, 2025, the DeepSeek day. Yahoo also counts 18 single-session drops of $200 billion or more by US stocks in 2026 through Wednesday, against 13 in all of 2025 and none before 2022. A chart in the same article shows 19 through October 29. By Yahoo's count, Alphabet is down about $1.4 trillion from its May record. yahoo

The numbers behind the drop are consistent across sources. Search and other revenue rose 12% to $63.35 billion. Fortune says analysts expected 15% growth, after 17% in Q2 and 19% in Q1. Free cash flow was negative $9.07 billion, compared with the $2.7 billion analysts expected, according to CNBC. Fortune puts quarterly capex at $51.4 billion, more than twice the prior year. CNBC says full-year capex guidance stayed at $195 billion to $205 billion, and Fortune reports guidance of about $200 billion for 2026 and $300 billion for 2027. The rest of the business did well. Fortune reports Google Cloud revenue up 103% to $30.71 billion, with an operating margin of 36.2% against 23.7% a year earlier, and Yahoo reports the Gemini app passing 1 billion users. Fortune also notes a one-time $8.7 billion mark-to-market loss on Alphabet's SpaceX stake. CNBC reports that chief business officer Philipp Schindler blamed part of the search shortfall on weaker monetization in AI Mode, and said ads in AI Overviews and AI Mode are now "scaling," including internationally. yahoo

The contrast that matters most is in Fortune. Pichai said query growth is at a five-year high, while BofA's Justin Post and Nitin Bansal wrote that search growth slowed "despite accelerating y/y query growth." In other words, usage is up and revenue per query is down. Fortune also cites a Pew finding from July 2025: users clicked a traditional result on about 8% of visits that showed an AI summary, against 15% without one. That study is more than a year old and was not re-checked here. BofA kept its Buy rating, cut its price target to $380 from $430, cut its target multiple to 24x from 27x, and said Google search now trades at roughly 10x earnings, below the 13x of the energy sector. yahoo

The rest of the week, per CNBC's Friday report: Amazon kept its 2026 capex forecast at $220 billion and posted a second straight quarter of negative free cash flow (CNBC). Yahoo adds that AWS grew 39%, its fastest pace in nearly five years, and that Amazon's trailing twelve-month free cash flow was a record outflow of more than $42 billion. Intel guided Q4 revenue of $15 billion to $16 billion and adjusted EPS of 25 cents, against LSEG consensus of $17.1 billion and 43 cents. It also said an explosion at its Israel fab, which killed two workers, will cost about $600 million in lost revenue over two quarters. CNBC says the US government's 10% stake lost about $12 billion but remains profitable, and Nvidia's stake lost about $6 billion. The Nasdaq fell 2.9%, Nvidia 4%, Broadcom more than 5% and Qualcomm 11%. Microsoft was the exception. CNBC says its quarterly capex and finance leases were $45.2 billion, below the $50.67 billion analysts expected, and its free cash flow fell 27% to $18.63 billion but stayed positive. Its shares were roughly flat. yahoo

Facts vs. opinions. The reported figures are facts: revenue, free cash flow, capex, guidance and share moves. Even so, the market-value loss is given three different ways ($475 billion, about $482 billion, about $500 billion), and Intel's fall appears as 28.58% (Morning Brew, Yahoo's sidebar) and 29% (CNBC). Calling this "the first quarter that AI is starting to seriously hurt their search business" is an opinion. Pivotal Research's Jeffrey Wlodarczak attributes it to "the bears," not to himself. Morning Brew's line that Intel fell because of "weak PC demand amid high AI spending" is a summary. Other outlets link weak PC demand to AI data centres buying up memory, but those reports were not opened for this issue. The idea that AI Mode earns less per query than classic search is supported indirectly by Schindler's remark and BofA's note. Alphabet has not published a per-query revenue figure in anything read here. yahoo

Other perspectives. Analysts split clearly. Ben Barringer of Quilter Cheviot called it "a fairly messy set of results and a poorly handled earnings call," said Google "faces an innovator's dilemma," and said "Google has lost its edge with Gemini and is behind peers on agentic AI." Scott Devitt of Rosenblatt said "I don't think it's a sea change" and that "Alphabet, with its 4 billion search customers that become AI customers, is in an enviable position" (both via CNBC). CFRA's Angelo Zino told Yahoo the selloff "probably represents a buying opportunity." BofA sees tough comparisons through 2Q'27, "which could make Google a 2H'27 story," but also "potential for a growth uptick as early as 1Q." On Intel, Raymond James's Matt Orton said "I don't think this is a moment of reckoning but I think it's a moment where investors should take a long thought." yahoo

Why it matters. For an AI engineer, the useful signal is not the share price. It is that the companies buying the most inference hardware now report the money going out as clearly as the revenue coming in. Google Cloud more than doubling is direct evidence of demand for compute. Search slowing while queries rise suggests that the AI interface Google is building can earn less per query than the one it replaces, at least for now. Two things are worth watching. One is whether ad monetization in AI Mode closes the gap BofA describes. The other is whether a $300 billion capex plan for 2027, if Fortune's figure holds, survives a second quarter like this. Model pricing, rate limits and free tiers on Gemini and AWS ultimately depend on those budgets. The cash-flow pressure on Amazon and Alphabet makes aggressive price competition less certain than it looked in the spring. yahoo

Binance loses a record $6.3 billion: a DevOps laptop, a service container and a hot-wallet signer

1 source, 5 min read

What happened. On October 29, attackers took about $6.3 billion from Binance, most of it in roughly 20 minutes. It is the largest crypto theft on record. Binance suspended withdrawals until they pass what its co-CEO called "rigorous testing," which Morning Brew says should finish this weekend. Binance and its partners froze about $3 billion, and Binance says it will cover all losses. Binance said the attackers may have got in through an engineer's laptop. The FBI is investigating, and blockchain analysts point to the North Korean group blamed for the Bybit theft. morningbrew

Details. The best technical account so far comes from TRM Labs (TRM Labs). TRM says about USD 6.36 billion left Binance wallets on ten blockchains, starting at 20:05 UTC, valued at the price when taken. At current prices it is about $6.23 billion. A second wave followed at about 23:05 UTC. Ethereum accounted for about $3.79 billion: 200,630 ETH plus 185 other tokens, including 1.875 billion USDT, 823 million USDC and 88.5 million RLUSD. Bitcoin accounted for 18,270 BTC, about $1.47 billion. TRM's central judgement is: "In TRM's assessment, this points to a compromise of Binance's hot wallet signing infrastructure." TRM thinks the attacker could "make Binance's hot wallet signing system sign transactions of its choosing across chains." morningbrew

TRM shows the attack was planned. "Preparation began at least 36 hours before the theft." The attacker's operator wallet was funded with three small withdrawals from Tornado Cash starting the morning of October 28. The attacker apparently rehearsed, and "Four hours after that rehearsal, the same code was installed on a Binance hot wallet." On Ethereum, 50 pre-signed EIP-7702 authorizations were submitted at 20:20 UTC. EIP-7702, which arrived with the Pectra upgrade in May 2025, lets an ordinary account delegate its execution to contract code. The main sweep came 3 minutes 24 seconds after delegation and took 167,464 ETH plus 63 tokens, including 1.615 billion USDT. TRM says all signatures were produced within about half an hour. It lists "signatures obtained in advance then executed in one burst" as a trait shared with earlier North Korean heists. morningbrew

Binance's own account, via BleepingComputer, is thinner (BleepingComputer). A DevOps engineer's laptop was compromised, by a method not stated. From there the attacker reached a service container and submitted unauthorized withdrawal transactions from inside it. Co-CEO Richard Teng: "To be clear: the attacker did not simply use one engineer's permissions, and this was not a single 'one-click' action." Binance says it "uncovered additional points of breach" and has found no evidence that a wallet private key leaked. TRM notes: "Binance has yet to publish a root cause." morningbrew

The freezes were fast. Per TRM, Tether froze 971.1 million USDT at 21:01 UTC, 37 minutes after the main sweep. Circle froze 823 million USDC at 21:04 UTC. Ripple froze 40.5 million RLUSD, and World Liberty Financial froze 27.8 million USD1. The Arbitrum Security Council moved 8,426 ETH and 74.6 million ARB. TRM counts about $2.7 billion frozen as of 00:30 UTC on October 30, and notes Binance's own figure of $2.54 billion verified as of 23:35 UTC. Morning Brew and CNBC report Binance's later estimate of about $3 billion. The attacker still holds more than 420,000 ETH and 18,270 BTC, about $2.4 billion that no issuer can freeze. The laundering was costly. TRM says 325 million USDT was sold for about 122,300 ETH at an implied price of about $2,650 per ETH, while ETH traded at $2,300 to $2,350, a discount of more than 10%. Another 60 million USDT was supplied to Aave to borrow about 19,760 ETH. morningbrew

Facts vs. opinions. The flow of funds on-chain is public and has been reconstructed independently, so it is the firmest part of the story. Attribution is an assessment. TRM names TraderTraitor based on what it calls "robust on and off-chain intelligence sources," and BleepingComputer says Chainalysis agrees "with a high degree of confidence." Binance has not attributed the attack to anyone, and Teng said "Our investigation has not attributed responsibility to any individual." The laptop-to-container-to-signer path combines Binance's partial disclosure with TRM's inference, and no root-cause report exists. Some figures in Morning Brew were not confirmed elsewhere: that the theft is almost twice all 2025 crypto thefts combined, and that Binance's cyberattack reserve held only $1.23 billion in bitcoin. The totals also moved as the story developed. BleepingComputer once headlined $5.8 billion, TRM says USD 6.36 billion, and Chainalysis, per BleepingComputer, says $6.4 billion. morningbrew

Other perspectives. Investigators describe a race against time. Chainalysis CEO Jonathan Levin told CNBC he saw "laundering patterns we have not seen before" and warned that "As you get further and further away, the amount of seizure opportunities go down" (CNBC). TRM's Ari Redbord called it "It's this continued cat and mouse game where bad actors are trying to obfuscate their transactions." Joe Hancock of Mishcon de Reya said of the quick stablecoin swaps: "The attackers knew that route would be closed to them and so got out ahead of that." CNBC found no expert commenting directly on Binance's custody design. morningbrew

Why it matters. Set aside the crypto, and this is a familiar infrastructure failure at very large scale. What follows is analysis, not reporting. If TRM is right that the signing system signed whatever it was handed, then the signer trusted its caller. A service container could request signatures, and the signer had no independent controls on how fast or where money could go. A well-designed signing service enforces its own policy: withdrawal velocity limits, destination allowlists and human approval above a threshold, checked inside the signer rather than by the service calling it. Fifty pre-signed EIP-7702 delegations are especially telling. A delegation can turn a hot wallet's whole balance into something attacker code can spend, and a policy engine that only checks individual transfer amounts may never see it as a withdrawal. The other lesson is that the effective backstop was centralised. Stablecoin issuers froze about $2.67 billion within hours, and the unfreezable part, ETH and BTC, is what got away. Anyone running agent systems with signing or payment rights should take note. The authority to spend should sit behind its own policy check, not behind whichever process happens to hold the credential. morningbrew

A-RSI: when should a model that trains itself get new tasks?

1 source, 4 min read

What happened. A 20-author team posted A-RSI, "Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement," on arXiv on October 30. The authors list Chinese, European and US affiliations, including the Max Planck Institute for Intelligent Systems (MPI-IS) and Carnegie Mellon (CMU) on the paper page. "Zero-data" means the model trains on tasks it generates itself, with no external training data or human-written labels. The question the paper asks is not what tasks to generate but when the current pool of training tasks has stopped being useful. The method watches reward variation and accuracy in the training logs, never validation or test scores, and swaps in a fresh pool when the signal runs out. It was the most-upvoted lead paper on Hugging Face's daily list, at 150 votes (Hugging Face paper page). huggingface

Details. The tasks come from a static library of programs written by a model. The programs generate multimodal questions and compute their answers, so labels come from code rather than majority voting. That matters because the authors argue majority-vote labels were a key weakness in earlier work. Training uses GRPO, a reinforcement-learning method that compares groups of sampled answers, in windows of 100 updates, up to eight windows. The old pool has 500 tasks, and each refresh replaces it with 400. The refresh detector has two thresholds: 85% signal retention and a ten-point rise in accuracy. The signal is the number of prompts that still produce a non-zero advantage, meaning prompts where sampled answers still differ in reward and so still teach the model something. On Qwen's original pool, that number fell from 239 to 195, a ratio of .816, which triggered a refresh. Training accuracy on the first fresh pool was .35, against .57 at the end of the old pool, so the fresh tasks were harder. huggingface

Paper facts. arXiv preprint 2610.38568, v1, dated October 30, 2026. No venue is stated. The authors are led by Yu Liu, Youxing Li and Yuedong Tan (equal contribution), with corresponding authors Cong Cao and Pei Fu. Co-authors include Radu Timofte, Nicu Sebe, Barnabas Poczos and Jian Luan. The main model is Qwen3-VL-8B-Instruct, evaluated as the average accuracy across eleven benchmarks (the paper calls it Macro11) with an 8,192-token answer limit. The average rises "from 57.77 to 61.05 with 800 fresh tasks." Baselines: No-Refresh scored 58.48 at window 4 and 58.35 at window 6, then stopped at a later checkpoint. At the final checkpoint, refreshing continuously scored 60.64 using 3,200 new tasks, refreshing periodically scored 59.89 with 1,200, and refreshing from a difficulty frontier scored 59.59 with 800. Gain per 1,000 new tasks is 4.09 for A-RSI, against 0.90, 1.77 and 2.28 for those three. On the second model, MiMo-VL-7B-SFT, the score goes from 59.08 to 60.71, against 60.01 for continued reuse. A-RSI is only +0.02 over refreshing every window and +0.18 over periodic refresh, which is effectively a tie. An LLM judge, Qwen2.5-14B-Instruct, is used for grading. Ablations against a random-selection control at 60.29 show every "clever" selection rule doing worse. Prioritising failures scored 57.67, a "learnable band" of difficulty scored 58.39, and always-wrong tasks scored 57.89. Code release: none found in the page text. The authors say their prompt and checker follow "the public MM-Zero evaluation code." huggingface

Facts vs. opinions. The headline gain of 57.77 to 61.05 is a single run on one model family and eleven benchmarks. The authors themselves say single runs and small margins "do not establish an advantage of detector-triggered timing." The more defensible finding is narrower: fresh random task pools beat reusing stale ones (+2.05 on Qwen, +0.70 on MiMo), and simple random selection beat each of the targeted selection rules tested. The admission that the generator's task families and 19 thematic terms were "informed by the evaluation benchmarks" is a real contamination risk, because a self-improvement loop built partly around the test themes is not fully zero-data in spirit. huggingface

Other perspectives. The most useful part of the paper may be its reproduction of MM-Zero, a prior zero-data multimodal self-improvement system. The authors reproduce its published numbers closely: base 50.27 against a reported 50.03, and v5 54.25 against 54.18. They then argue that most of MM-Zero's gain comes from answers that finish within the token limit, not from better reasoning. An untrained base model at a 16,384-token limit scores 51.12 on four datasets, against MM-Zero's 51.08 at 4,096. Truncation at 4,096 tokens is 2.84% for MM-Zero v5 against 20.1% for the base model. On 4,006 base items that never finished, accuracy went from 0.30 to 29.18 after MM-Zero training. They also report that MM-Zero's majority-vote label agrees with the generator's own answer only 5 to 12% of the time, and that about 16% of its released images are nearly black. No response from the MM-Zero authors was found. huggingface

Why it matters. For anyone tracking recursive self-improvement, the paper puts a cap on the gains and recommends a reporting standard. The cap: with careful auditing, a zero-data loop on an 8B multimodal model gains about three points across eleven benchmarks, and even that is a single run. The standard: zero-data papers should report the base model at relaxed token limits, truncation rates, and gains at several limits. That applies to text-only self-play and to reasoning RL generally, because "the model learned to stop rambling" is easy to mistake for "the model learned to reason." The finding that random selection beat failure-prioritised or band-limited curricula is also worth remembering before building a clever curriculum into an agent's training loop. The detector itself, which watches the share of prompts that still produce learning signal, is cheap to put in any GRPO pipeline as a warning that a task pool has gone stale. huggingface

One wide search beat a search agent: a small Bedrock RAG experiment

1 source, 4 min read

What happened. In a guest post for Decoding AI, Marcelo Acosta Cavalero built two versions of a support chatbot on Amazon Bedrock for a fictional outdoor store. RAG (retrieval-augmented generation) means the app searches a document store and the model writes its answer from the passages returned. One version ran a single wider search for nine passages. The other was an agent that could search again based on what it had read. His conclusion: "For this chatbot, I'd keep the single search with nine passages." The agent cost more, took longer and passed fewer answers. source

Details. The setup is small and fully described. There are 24 policy articles of about 29 KB in total, stored in S3 and split by Bedrock Knowledge Bases into passages of about 200 tokens. Titan Text Embeddings V2 converts passages to vectors, stored in S3 Vectors, and GLM 5 on Bedrock writes the answers. Nine passages was chosen as "the smallest setting tested that retrieved all the required policy text" on eight practice questions. Five versions were compared on 24 customer questions (eight direct lookups, eight that combine policies, eight asking for information not in the documents), for 120 answers. The five were: three passages, nine passages, the agent, three passages followed by a self-check, and every document sent in the prompt. An AI evaluator graded each answer against rules written before the run, and code checked citation IDs and recorded costs and times. Headline: "The agent cost 23% more than the nine-passage search and took 72% longer, while producing two fewer passing answers." Sending every document "matched the nine-passage score at almost three times the model cost." The agent hit its search limit on nine questions, and "ten of its 25 extra searches went to questions whose answers weren't in the documents." The full table of pass counts appears to have been an image. The text gives only the differences, so absolute pass rates are not stated. source

The two failure cases are the most useful part. In the first, a 60-day trial on hiking boots starts when an order is collected in store. The nine-passage search found that rule and applied it. The agent found it too, then named both delivery and collection as possible start dates and said it couldn't tell which applied. In the second, a customer paid USD 120 for boots plus USD 6 delivery, and the return label costs USD 7, so the correct refund is USD 113. The three-passage answer got it right, and so did the self-checked version. The nine-passage, agent and all-documents versions all said USD 107, subtracting the separately paid delivery charge again. The author suspects a duplicate statement that delivery charges are nonrefundable encouraged the extra subtraction, and says removing that passage and re-running would test the idea. source

Facts vs. opinions. The measured facts: one model, one corpus, 24 questions, one run each, an LLM judge, and costs counting model tokens only ("storage, embeddings, and vector operations are separate"). A difference of "two fewer passing answers" out of 24 is not statistically meaningful on its own, and the author does not claim it is. His conclusions are explicitly provisional. Before release he would "test new customer questions, run them repeatedly, and have someone who knows the policies check a sample of the judgments." The article says AWS's guidance recommends standard retrieval for direct lookups and agentic retrieval for more involved questions. That AWS page was not opened here, so the summary is credited to the article. source

Other perspectives. The cost-and-latency result matches what Co-Design Bench, the next story, finds at much larger scale: more agent turns and tokens do not reliably buy more completed work. The opposite case is easy to make, and the author makes it himself. An agent helps "when the next useful search depends on what the previous result reveals," such as following a product fault from a component to a service bulletin to affected versions. The test corpus is 29 KB and fits entirely in a prompt, which is close to the worst case for an agent, because one wide search can surface almost everything. On a corpus a thousand times larger, nine passages would not cover a multi-part question and the result could reverse. source

Why it matters. Three practical points carry over. First, the diagnostic rule: "when an answer fails, look at the passages it received. If the policy was missing, work on retrieval. If the policy was present, work on how the answer applies it." That splits evaluation into a retrieval part and an application part, which managed tools can track separately. The article names Bedrock's bring-your-own-responses RAG evaluation for that. Second, more context can actively hurt. The refund error appeared only once more text was included, so passage count needs tuning like any other parameter. Third, money arithmetic should not be left to the model. The author suggests asking for an explicit breakdown, or better, having the chatbot request a quote from an existing returns service, so that "nobody would be coding 'this answer is 113.'" Routing only hard questions to an agent is offered as a hypothesis to test, not a result. source

Co-Design Bench: real users' multimodal coding requests, and rankings that depend on the harness

1 source, 4 min read

What happened. A team from Alibaba's Token Hub, Fudan, Zhejiang University, CUHK, SJTU and NTU released Co-Design Bench, a "reference-free benchmark of 273 tasks curated from 37,463 deduplicated queries across 26,314 real user coding sessions." Reference-free means grading does not compare against gold code or target screenshots. The tasks ask coding agents to turn users' multimodal requests into working revisions of running apps. Its headline: the best of ten endpoints "passes an average of 69.9% of criteria per task but fully satisfies only 26.4% of tasks" (Hugging Face paper page). huggingface

Details. There are two tracks. Multimodal Editing (ME) has 100 localized edits, 80% of which come with attachments: 32 audio, 27 video, 17 document and 4 image tasks, with 20 having none. Repository Editing (RE) has 173 coordinated revisions, 98.8% of them React apps. In total there are 2,331 hidden acceptance criteria written by experts. Grading uses what the authors call judge-then-refute. A judge grades each criterion from source code and browser evidence, a separate LLM verifier challenges positive verdicts, and visual and interaction claims are audited. Runs are capped at 3 hours and 120 turns. The full grid is 2,730 model-task pairs and 135,370 tool calls. The leader on strict completion (criteria passed per task) is gpt-6-astra at 0.699, followed by opus-5 at 0.661, qwen3.8-max-0902 at 0.653 and the open-weights Qwen3.8-27B at 0.644. On RE, Qwen3.8-27B scores 0.645 against opus-5's 0.642. Full-pass rates are low across the board. The leader fully satisfies 41.0% of ME tasks and 17.9% of RE tasks, so 73.6% of tasks keep at least one unmet criterion. huggingface

Paper facts. arXiv preprint 2610.35138, v1, dated October 27, 2026. No venue is stated. First authors are Xilin Wei, Shengyuan Ding and Xinyu Fang; corresponding authors are Dahua Lin, Hang Zhang and Shuai Bai. Ten endpoints were evaluated under one shared agentic harness. Robustness checks: rescoring with three alternative judges shifts graded scores by at most 0.053, and the ordering holds. Two experts reviewed 200 outputs. The project page links a GitHub repo for "Code & evaluation" and a Hugging Face dataset listed as "96 ME tasks · 521 criteria" (project page). That is only part of the 273-task, 2,331-criterion benchmark. A release of the 173 RE tasks is not mentioned, and no license is stated. Stated limitations beyond these are not covered in the text read. huggingface

Facts vs. opinions. The harness-swap result matters more than the leaderboard. When the two compared models, kimi-k3 and Qwen3.8-27B, were each run under three other agent harnesses (codex 0.142.0, opencode 1.18.13, qwen-code 0.21.11), the harness rankings inverted completely between the two models on ME (ρ = −1.00, a Spearman rank correlation). On RE the correlation was +0.50. The gap between the two models ranged from +0.027 to +0.127 on ME and from −0.040 to +0.122 on RE depending on harness. In plain terms, the choice of harness can move a model's score by as much as the gap between models, and on RE it can even reverse which model wins. The headline ranking is a fact about one shared harness. It is not a fact about the models in general. Also, the curators come from Alibaba's Qwen organisation, and Qwen models do well. Nothing read here suggests bias, but the conflict should be noted. huggingface

Other perspectives. The benchmark has interesting data on effort. On ME, gpt-6-astra used 32 turns and 1.24 million tokens per case for a strict score of 0.754. GLM-5.2 used 69 turns and 7.29 million tokens for 0.572. On RE, Qwen3.8-27B used 57 turns and 6.72 million tokens for 0.645, against gpt-6-astra's 48 turns and 2.33 million tokens for 0.667. A cheap per-token price can therefore be eaten up by token-hungry behaviour. The paper also motivates the agentic setup with an earlier test on 150 tasks (WebCompass-edit): qwen3.7-plus rose from 73.6% feature completion in a single attempt to 95.1% when it could act as an agent. So agent loops clearly help on narrow edits, yet on real users' composite requests the leader fully delivers about one task in four. The breadth result points the same way. RE edits spanning 3 to 4 change types score 0.591 against 0.654 for 1 to 2 types, and zero-pass rates rise from 0.9% to 5.8%. huggingface

Why it matters. This connects to the Bedrock story. More steps do not mean more correct outcomes, and how you evaluate matters as much as the scores. For product teams, criteria derived from real sessions ("did the edit do what the user asked, verified in a browser") are closer to user satisfaction than unit tests. The gap between per-criterion completion (69.9%) and full completion (26.4%) is the gap between "mostly works" and "done." For anyone picking a coding agent, the ρ = −1.00 inversion means benchmark rankings should be checked under your own harness before any decision. The media tool data is a design hint too. OpenCV appeared in 68.5% of video runs and ffmpeg in 31.5% of video and 25.0% of audio runs, so agents are improvising media pipelines that a harness could supply directly. huggingface

From the archive

Anthropic's "dreaming" memory and multi-agent orchestration for Claude (AlphaSignal, May 7, 2026)

2 sources, 2 min read

What it said. AlphaSignal's item, six months old, said Anthropic was adding self-learning memory and multi-agent teamwork to Claude, framed as "Capabilities your team couldn't ship before." This is an old item revisited, not news. alphasignal

How it aged. The features are real, but after six months they are still not generally available. Anthropic's blog post, "New in Claude Managed Agents" (the page shows May 19, 2026, later than the AlphaSignal item, so it may have been updated), describes three things (Claude blog). Dreaming "reviews agent sessions and memory stores, extracts patterns, and curates memories so agents improve over time." Outcomes lets "you write a rubric describing what success looks like and the agent works toward it," with "A separate grader" in its own context window. Multiagent orchestration "lets a lead agent break the job into pieces and delegate each one to a specialist with its own model, prompt, and tools." The evidence offered was vendor and customer claims. Harvey: "Completion rates went up ~6x in their tests." Wisedocs: "Reviews now run 50% faster." Outcomes "improved task success by up to 10 points over a standard prompting loop," with "+8.4% task success on docx and +10.1% on pptx." None of these come with published methodology in the material read. alphasignal

As of November 1, Anthropic's current documentation still marks dreams as "research-preview," with access by request (Claude docs: Dreams). The model list has grown to Claude Fable 5.1, Fable 5, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5, Opus 4.8, Opus 4.7 and Sonnet 4.6, so the feature has been carried through two model generations without graduating. The docs make the mechanism clearer than the launch did. Agents' memory writes are "local and incremental," so "over many sessions a memory store accumulates duplicates, contradictions, and stale entries." A dream is an asynchronous job that reads a memory store plus 1 to 100 session transcripts and produces a new store. "The input store is never modified, so you can review the output and discard it." Dreams "typically take minutes to a few hours," are billed at standard token rates, and cost "scales roughly linearly with the number and length of input sessions." They are not eligible for zero data retention. The docs also say memory store calls now use a different beta header, agent-memory-2026-07-22, which suggests the memory API was revised in July. That is an inference from the header date. alphasignal

Why it matters now. The design has held up better than the hype. "Self-learning memory" turned out to be an offline consolidation job that produces a reviewable copy, not a model that updates its own weights or silently rewrites its memory. That is the right engineering choice: a versioned memory store, offline batch consolidation, and human review before promotion. It is also directly relevant to the agent-session-memory problem Decoding AI's series raises ("Why Does Every Agent Session Start From Zero?" appears in its index). The open questions are the ones the launch skipped. How do you evaluate a dreamed memory store before promoting it? How large is the ~6x Harvey effect outside one customer's tests? And what does it mean for portability to keep agent memory in a vendor-hosted store that is not zero-data-retention eligible? A research preview lasting six months is also a signal in itself. alphasignal +1

Qwen's FlashQLA linear-attention kernels (AINews, April 30, 2026)

2 sources, 2 min read

What it said. AINews led its kernels section on April 30 with Qwen's FlashQLA. These are linear-attention kernels (attention whose cost grows linearly rather than quadratically with sequence length) written in TileLang, with reported 2–3× forward and 2× backward speedups, "especially for small models, long-context workloads, and tensor-parallel setups." The design relied on what Qwen called gate-driven automatic intra-card context parallelism (CP), algebraic reformulation, and fused warp-specialized kernels. It was "explicitly positioned for agentic AI on personal devices." The same roundup covered vLLM's Blackwell throughput, a torch.compile walkthrough, John Carmack's note on a 10× regression in torch.linalg.solve_ex between 511×511 and 512×512, and Zhipu's GLM-5 serving postmortem. Those items were not re-checked here. This is an archive item. x

How it aged. FlashQLA is maintained, modestly adopted, and its most important adoption step is indirect. As of November 1 the GitHub repo shows 744 stars and 85 forks, and it is MIT-licensed (GitHub: QwenLM/FlashQLA). The README's speed claim is unchanged: "achieving 2-3× forward speedup and 2× backward speedup over the FLA Triton kernel." The baselines are pinned at FLA 0.5.0, Triton 3.5.1, FlashInfer 0.6.9 and TileLang 0.1.8, with head configurations "h_k,v ∈ {64, 48, 32, 24, 16, 8}, corresponding to TP1 through TP8." The README's news section lists v0.1.3 in September 2026. The most consequential change is that since v0.1.2 FlashQLA "serves as a backend for flash-linear-attention's GDN," the Gated DeltaNet layer used in Qwen's hybrid models. Anyone using the standard flash-linear-attention API can get the speedup without adopting a new library. vLLM and SGLang are not mentioned on the repo page, so direct integration into the major serving engines is not shown. A small discrepancy: the benchmark install command pins flashinfer-python 0.6.13, while the stated baseline is 0.6.9. x

The "personal devices" framing is weaker than the launch suggested. The README's detailed benchmark files are for H200 and GB200, which are datacentre parts. The concrete numbers live in a chart image and linked text files, so no independent reproduction was found. The 2–3× figure remains a vendor benchmark against a Triton baseline. Context from today's issue: Qwen3.8-27B is competitive with frontier closed models on Co-Design Bench's RE track (0.645 against opus-5's 0.642). That makes efficient kernels for mid-size hybrid-attention models more relevant to local and on-premises agents now than in April. x +1

Why it matters now. Kernel releases rarely get adopted as standalone libraries. They succeed by becoming the default backend inside something people already use, and FlashQLA's path through flash-linear-attention is a typical example. For anyone running Qwen-family hybrid models locally, check whether your inference stack calls flash-linear-attention's GDN path. If it does, FlashQLA may already apply or be one dependency away. The open question from April is still open: is there an independent, edge-device benchmark (consumer GPUs, small batch, long context) supporting the "agentic AI on your personal devices" positioning? None was found as of November 1. x

Meta's unified AI agents for capacity efficiency (Engineering at Meta, April 16, 2026; via TLDR DevOps, May 4)

2 sources, 2 min read

What it said. Tommy Tran and Michael Zetune described how Meta's Capacity Efficiency Program uses one agent platform for two jobs. "Offense" finds and ships proactive optimizations. "Defense" detects and mitigates production regressions (Engineering at Meta). The architecture has two layers. MCP tools (single-purpose interfaces for querying profiling data, experiment results, config history, code and docs) sit underneath. Skills, meaning encoded senior-engineer expertise, sit on top, such as "look for recent schema changes if the affected function handles serialization." The key line: "The same tools can power both offense and defense. Only the skills differ." On defense, Meta's regression detector FBDetect "catches thousands of regressions weekly," down to "regressions as small as 0.005%." Traditional correlation identifies the culprit pull request, and an agent drafts a fix-forward pull request for the original author. On offense, the agent produces a guarded candidate fix that appears in the engineer's editor for one-click application. The program claims "hundreds of megawatts (MW) of power" recovered, and automated diagnosis that turns "~10 hours of manual investigation into ~30 minutes." This is a six-month-old item. fb

How it aged. No Meta follow-up to this specific post was found, so there is no later measurement of the claims. Rereading it with that in mind, the original never quantified what the AI itself did. The megawatts belong to the whole program, and the AI systems "contribute to supporting this effort." The post gives no figures for pull-request acceptance rate, fix accuracy or failures. Root-causing still relies mainly on non-AI correlation, and the agent's fix-forward pull request is sent to the original author for review. The only adoption evidence is "Many engineers at Meta use our efficiency tools." The adjacent Meta work that did develop is KernelEvolve, an agentic kernel-authoring system. Its arXiv technical report was revised to v4 on July 6, 2026, and the abstract claims "100% pass rate on all 250 problems" of KernelBench, "100% correctness" on "160 PyTorch ATen operators across three heterogeneous hardware platforms," and development time cut "from weeks to hours" (arXiv: KernelEvolve). The arXiv page lists no venue. Search results point to an ISCA 2026 listing, which was not opened here and is therefore not confirmed. fb

Why it matters now. The tools-plus-skills split has become standard practice in agent engineering: a shared tool layer with narrow, auditable interfaces, and many small expertise documents on top. Meta's post is one of the clearer hyperscale examples, and its central claim, that one tool layer serves many workflows with "few to no new data integrations," is plausible and cheap to copy. The lesson from rereading it is about evidence. Infrastructure-agent posts tend to report program totals (megawatts) and best-case time savings (10 hours to 30 minutes) rather than per-PR success rates. Co-Design Bench's gap between 69.9% of criteria passed and 26.4% of tasks fully done is a reminder of how much those two kinds of number can differ. Anyone building a similar system should report acceptance and revert rates from the start. fb +1

"Mitigating Metrics Malaise: Claude on Goodhart's Law" (Eating Policy, April 22, 2026; via TLDR Product, April 24)

5 sources, 2 min read

What it said. Jennifer Pahlka, writing on her Substack Eating Policy, describes a long-running friendly argument with Dave Guarino, who "thinks getting the metrics right is one of the biggest levers for making a system deliver better results." She agrees in principle but has seen good measures go bad too often, since Goodhart's Law "is always waiting in the wings." After a discussion of Justin Fanelli's use of metrics in the Navy, she decided "I should spend more brain cells on mitigations for Goodhart's Law instead of fighting Dave," and asked Claude (Eating Policy). Claude's answer, which both she and Guarino found persuasive, calls Goodhart's Law "one of the deepest traps in organizational life" and concludes "the problem is structural — that measurement and reality inevitably diverge." It lists six habits. Leaders "treat metrics as evidence, not truth." They rotate and retire measures, invest in qualitative feedback, make it safe to report gaming, hold several measures in tension so that gaming one costs another, and make the purpose clear so people notice when a metric points the wrong way. It also names a failure mode: responding to metric breakdown with more metrics or tighter accountability. A commenter, Edith Hughes, adds that triangulating across five data types (surveys, unprompted feedback, transactional, operational and employee data) reveals the "WHY" behind a single metric's "WHAT." This is an archive item. eatingpolicy

How it aged. It is an essay, so "aging" here means whether later evidence supports it. No later follow-up by Pahlka was searched for. Today's issue, though, has four fresh examples of a metric drifting from reality. A-RSI finds that a zero-data system's benchmark gains were largely about answers finishing under a token cap, because a metric that drops truncated responses rewarded length control rather than reasoning. The authors' fix matches habit five: report several measures (several token limits plus truncation rates) together. Co-Design Bench shows a model ranking that inverts when the harness changes, so the leaderboard is "evidence, not truth" in a very literal sense. Its two headline numbers, 69.9% of criteria against 26.4% of tasks fully done, are measures held in tension. Alphabet reported query growth at a five-year high while search revenue growth slowed, the case where a usage metric looks good and the business underneath is worse. And the Bedrock test's advice, look at the passages each failed answer received, is habit three, qualitative contact with reality, applied to evals. eatingpolicy +4

Why it matters now. The list holds up because it treats Goodhart's Law as structural rather than as a moral failing, which is the right frame for AI evaluation too. Two caveats. First, the post is an LLM's synthesis of management literature with no citations. Its claims about what "great leaders" do are plausible generalizations, not findings. Second, rotating metrics has a cost the post does not mention: you lose comparability over time, which matters for tracking a model or a product across versions. A practical way to apply it to eval suites is to keep a stable core benchmark for continuity, rotate a held-out set regularly, always pair a completion metric with a cost or length metric, and give someone the job of reporting where the eval and production disagree. eatingpolicy

TMAS: scaling test-time compute through multi-agent synergy (Hugging Face papers, May 13, 2026)

2 sources, 3 min read

What it said. TMAS from IQuest Research, with one Beihang University author, treats inference (test-time compute) as a collaboration among five agent roles: solution, verification, summary, experience and guideline (Hugging Face paper page). Two memory stores support them. An experience bank holds verified intermediate results, skills and pitfalls. A guideline bank records strategies already tried so later attempts don't repeat them. Each round, a candidate either refines past attempts or, with probability ε, explores a new strategy. A hybrid-reward RL scheme built on GRPO adds an experience-utilization bonus and a novelty reward to the correctness reward. This is a revisited item from May. huggingface

Paper facts. arXiv preprint 2605.10344. The project's GitHub repository gives a May 11, 2026 release date; the Hugging Face page gives no date. No venue was found. Authors include Nan Jing and Qing Yi (equal contribution) and corresponding author Chuan Hao. The main benchmarks are IMO-AnswerBench-50 and HLE-Math-100, with AIME26 and HMMT-25-Nov in the appendix, where "differences between methods were small." Models are Qwen3-30B-A3B-Thinking-2507 and Qwen3-4B-Thinking-2507, with only the 4B model RL-trained (190 steps on 256 H20 GPUs). Inference uses N=8 parallel solutions, M=8 verifiers per solution, ε=0.2, up to 20 iterations and 128K maximum output tokens, scored by Pass@1. Baselines are Majority Vote@64, Self-Refine, Verify-Refine, PaCoRe and RSE. At iteration 19, the 30B model with TMAS scores 40.50 on IMO-AnswerBench-50 (best baseline RSE 38.00) and 35.38 on HLE-Math-100 (best baseline PaCoRe 32.78). The 4B model goes from 17.06 and 17.41 with TMAS alone to 30.88 and 28.16 with hybrid RL. With RL, the gap to the 30B model shrinks from 23.44 to 9.62 points on IMO and from 17.97 to 7.22 on HLE. The limitation, as it appears in search-result text and not re-checked on the arXiv page: TMAS was not evaluated on frontier models because of compute and API cost. huggingface

How it aged. Mostly by not shipping. The paper page links code at a GitHub repo. As of November 1 that repo has 6 stars, 0 forks, 4 commits and one open issue, and contains only a figures folder and a README (GitHub: TMAS-code). The README says: "Code, dataset and model checkpoints are currently under internal company review and will be released soon." The RL data file it describes (4,400 samples), the benchmark directory and the run scripts are not present. A search for citing papers or follow-ups found none. The related papers that came up (TUMIX, Team of Thoughts, ARTIS, FutureWeaver) were topical matches, mostly earlier work, not citations. That is a limited search, not a citation count. huggingface

Today's A-RSI paper gives a sharper way to reread TMAS. A-RSI shows that gains under a fixed token budget can largely reflect answers finishing rather than better reasoning. TMAS runs up to 20 iterations with 8 solutions, 8 verifiers each and a 128K output cap, against baselines like Majority Vote@64. The summary read here does not show a compute-matched comparison, such as equal total tokens across methods, so how much of the gain comes from orchestration and how much from simply spending more cannot be determined from what was read. The benchmarks are also small subsets (50 and 100 problems), where a couple of problems move scores by several points. huggingface +1

Why it matters now. The ideas themselves were sensible and have become common: a verified-experience memory, a tried-strategies list to avoid repeating attempts, and a novelty reward during RL. But TMAS is now an unreproducible preprint, because the code it advertised has been "under internal company review" for nearly six months. Treat its numbers as claims. If you want a multi-agent test-time scaling method to build on, prefer one with released code and compute-matched baselines. huggingface

GLM-5.2 vs Claude Opus 4.8: a one-shot WebGL platformer (Tech Stackups, June 18, 2026; via TLDR Dev, June 23)

2 sources, 3 min read

What it said. James Daniel Whitford of Ritza gave both models the same single prompt: build a 3D platformer in raw WebGL with no engine, using identical bundled CC0 assets, one attempt, and an instruction to verify before stopping (Tech Stackups). Opus 4.8 ran in Claude Code with high extended thinking. GLM-5.2 ran in the Pi agent via OpenRouter with high thinking; its higher Max tier was not used. Results: GLM-5.2 took 1h 10m 40s, 131,000 output tokens, 128 tool calls and $5.39 actually billed. Opus took 33m 30s, 216,809 output tokens, 153 tool calls and about $21.92, an estimate at list prices. Verdicts: "GLM-5.2 cost a fraction as much. Opus finished in half the time and shipped a cleaner game." GLM-5.2's "bugs were fundamentals," and Opus's "were edge cases, not broken basics." Because GLM-5.2 is text-only, it could not visually check its work, and "On a visual task, that was the difference between catching the rough edges and shipping them." The article cited Artificial Analysis's Intelligence Index v4.1 score of 51 for GLM-5.2, ahead of the next open-weights models (MiniMax-M3 and DeepSeek V4 Pro at 44, Kimi K2.6 at 43). Its conclusion: "Use GLM-5.2 when cost and openness matter and the work is mostly text and logic." This is an archive item. techstackups

How it aged. Three things changed. First, GLM-5.2 was superseded. Z.ai released GLM-5.3, and The New Stack's headline (the article body could not be read) says "Z.ai's GLM-5.3 goes open weight, but its new license aims at hyperscalers" (The New Stack). That headline matters for the article's strongest argument, that GLM-5.2 was MIT-licensed, "the rare frontier-adjacent model that no vendor can take away from you." If the successor ships under a more restrictive license, that argument is weaker for the model family going forward. Reports that GLM-5.3's weights were delayed for a safety review, and the specific license terms, appear in search snippets that were not opened. They are not confirmed here. A related New Stack headline on the same page, dated October 29, says "Mistral's Large 4 open weights arrive with a 1-point coding lead over GLM-5.3." techstackups

Second, today's Co-Design Bench gives a larger, multi-task comparison of the same models under one harness, and it broadly supports the article's verdict. On ME, glm-5.2's graded score is 0.629 against opus-4-8's 0.657. On RE it is 0.641 against 0.645. ME full-pass rates are 16.0% for both. That makes the two closer than the platformer suggested on multimodal editing and repository work combined. But glm-5.2 used 69 turns on ME against opus-4-8's 24, and the benchmark's own comparison puts glm-5.2 at 7.29 million tokens per ME case against gpt-6-astra's 1.24 million. The article's cost advantage was measured on a single run. On Co-Design Bench's agentic tasks, GLM's token appetite eats much of that per-token price advantage. Both have since been passed: opus-5 scores 0.749 on ME and glm-5.3-flash 0.599 on ME and 0.662 on RE. The article's point about visual self-verification also shows up there. ME tasks carry audio, video, image and document attachments that a text-only model must handle through tools. techstackups +1

Third, the article itself flagged that most of its benchmark table was Z.ai's model card plus Anthropic's self-reported figures. Its own correction (GLM-5.2's game did have a win condition, triggered only after all coins were collected) shows how fragile one-shot single-task comparisons are. techstackups

Why it matters now. The open-versus-closed gap on agentic coding is real but narrow on some tracks. On Co-Design Bench's RE track, the open-weights Qwen3.8-27B (0.645) roughly ties opus-5 (0.642). The decision rule should now be cost per completed task, not price per token: a model that is cheaper per token but takes three times the turns can cost more. Licensing, not weights availability, has become the real axis of "openness." Check the license on every new release instead of assuming it matches the last one. techstackups +1

Brain & consciousness

Reading path: State of Brain Emulation Report 2025, part 1: recording brain function (neural dynamics)

4 min read

What to read. The State of Brain Emulation Report 2025 is by Niccolò Zanichelli, Maximilian Schons, Isaak Freeman, Philip Shiu and Anton Arkhipov. It is an arXiv preprint (2510.15745, q-bio.NC), first submitted October 17, 2025, now at v3 from November 5, 2025, CC BY 4.0, with no journal reference (arXiv). It describes itself as a "comprehensive reassessment of the field's progress since Sandberg and Bostrom's 2008 Whole Brain Emulation roadmap," organized around three capabilities: "recording brain function (Neural Dynamics)," "mapping brain structure (Connectomics)," and "emulation and embodiment (Computational Neuroscience)." This step is the first of those. The report's website points to the full PDF, where the recording-specific detail starts at page 48 (report site). For orientation, read the website summary first, then the LessWrong launch post from January 28, 2026, which quotes the key numbers (LessWrong), then the PDF chapter. As a preprint, it is a synthesis and position document, not a peer-reviewed result.

Why this step comes first. The report itself states a framing that shapes everything after it: emulation means "reproducing the causal machinery," while simulation means "matching outputs." To reproduce causal machinery you need to know what the machinery does, which means activity data, not just wiring. The site states the bottleneck directly: "The main barrier to better brain emulation models is more and higher-quality experimental data." Not compute, not algorithms. Recording is where that data comes from, and it is the capability furthest from the goal.

Key ideas. First, the definition. The field "has not yet achieved whole-brain recording (≥ 95% of neurons and brain volume) at single-neuron resolution." The LessWrong post phrases the standard as "capturing more than 95 percent of neurons across 95 percent of brain volume simultaneously." By that bar, "no experiment to date has delivered that scale with single-neuron, single-spike resolution" in any organism. Second, the current state. The best results are "larval zebrafish with approximately 80% brain coverage," and C. elegans has "roughly 50% of nervous system neurons recorded at single-cell resolution." Calcium imaging in mice "now captures approximately one million cortical neurons in mice (though without resolving individual spikes)," while voltage imaging "resolves individual spikes in tens of thousands of neurons in larval zebrafish." The report summarizes two decades of progress as "about a two-order-of-magnitude increase in effective data bandwidth of neural recordings." Third, the trade-offs. Sampling is "typically well below neuronal firing rates (often 1-30 Hz for calcium imaging)," so calcium signals blur individual spikes. "Recording durations remain short (minutes to hours)," and "the need for head-fixation severely constrains behavior repertoires," meaning animals held still under a microscope can only do a narrow range of things. The LessWrong summary lists "significant trade-offs across spatial coverage, temporal resolution, recording duration, invasiveness, signal quality," and adds: "Even more challenging is recording of modulatory molecules like hormones and neuropeptides." Fourth, scale. The site lists C. elegans at about 300 neurons, zebrafish about 100K, fruit fly about 140K, mouse about 70M and human about 86B. A pull-quote on the site reads: "No foreseeable technology enables full neuron recording." That is about large brains, not small ones.

The strategic conclusion. The report's near-term target is sub-million-neuron animals: flies, larval fish, bees. The site says faithful capture is thought possible within roughly a decade at a cost in the low hundreds of millions of dollars, and the LessWrong post says "It seems plausible that this barrier will be overcome for sub-million neuron organisms in the upcoming years." The post also notes that perturbation tools like optogenetics, which switch neurons on and off with light, now make it feasible to propose "systematic reverse-engineering of neuron-level input-output relationships across entire small nervous systems." And it flags a mismatch: connectome reconstruction is outpacing activity recording, so methods to infer function from structure are needed. The site adds that the field is very small, fewer than 500 dedicated people, so a newcomer or funder can make a proportionally large difference.

Open questions to carry forward. First, how much activity data does emulation actually need once a connectome exists? If structure is cheaper than function, the decisive research may be in inferring dynamics from wiring plus sparse recordings. That links to the connectomics and computational-neuroscience parts of this reading path. Second, is calcium imaging at 1–30 Hz good enough to constrain models whose units fire faster, or is voltage imaging at whole-brain scale a hard prerequisite? Third, how should neuromodulators be handled, given the report calls them "even more challenging" to record? A model that captures fast electrical dynamics but not slow chemical state may match outputs without reproducing causal machinery, which by the report's own definition is simulation, not emulation. Fourth, can recordings in head-fixed animals over short sessions generalize to the behaviour of freely moving animals over long periods? The report's own limits suggest skepticism toward any near-term claim of a "whole-brain" mammalian emulation built mainly from activity data.

Quick hits

Research papers

8 sources, 1 min read

• Intrinsic Distribution Matching. The day's most-upvoted Hugging Face paper, at 238 votes. Only the title was available and the paper was not read, so its claims are not checked. huggingface
• Benchmarks as empirical measures over difficulty. "Your Benchmark Is an Empirical Measure Over Difficulty" (140 votes) argues from its title that a benchmark score reflects the difficulty mix of its items. That matches today's lesson from Co-Design Bench, where completion rates move with edit breadth. Not read beyond the title. huggingface
• Looking only when needed in diffusion VLMs. A certificate-guided visual routing scheme for efficient diffusion vision-language models (91 votes). Not read beyond the title. huggingface
• One encoder for parsing every modality. Youtu-Parsing-Omni proposes "One Encoder, One Schema, All Modalities" for document and multimodal parsing (55 votes). Not read beyond the title. huggingface
• Platonic Representation Hypothesis via JEPAs. A paper claiming to prove and improve the Platonic Representation Hypothesis, the idea that different models converge on similar representations, using joint-embedding predictive architectures (JEPAs) (53 votes). A "proof" claim from a title alone deserves scepticism until read. huggingface
• Real-time infinite world generation. L3G-World uses a latent 3D Gaussian memory for real-time, unbounded world generation (34 votes). Not read beyond the title. huggingface
• Retrieval heads in multimodal LLMs. "A Broader Look into MLLM's Retrieval Heads" (28 votes) extends work on the attention heads responsible for pulling facts from context. Possibly relevant to the Bedrock finding that extra context can corrupt an answer. Not read. huggingface
• Role-coupled tokenization for pixel-space generators. "Read, Allocate, Route" proposes a tokenization scheme for pixel-space generative transformers (26 votes). Not read beyond the title. huggingface

Agents & tooling

4 sources, 1 min read

• Professional intents into training tasks. WorkFactory turns professional task intents into executable work environments for scalable agent training (30 votes). It is relevant to the gap Co-Design Bench measures between narrow edits and composite real-world requests. Not read. huggingface
• Prompt-to-game with an adversarial agent team. VibeGame pairs an AI-native game engine with a "self-evolving adversarial agent team" (18 votes). The archive's GLM-5.2 versus Opus platformer test suggests visual self-verification is the hard part. Not read. huggingface
• "Errors as data" self-improvement for document parsing. A Towards AI post claims a recursive self-improvement loop helps OXR reach "SOTA accuracy and speed" in document parsing. Only the teaser was available. The state-of-the-art claim is unverified, and A-RSI's audit is a reason to ask for truncation and held-out numbers. medium
• Software engineering's "collapse." An opinion piece argues that "'It was never about coding' is a lie." Teaser only, no evidence to assess. medium

Local models & inference

4 sources, 1 min read

• Is local inference still worth it after Haiku 5.5? A Data Science Collective post's teaser claims "A cheap cloud model now matches Qwen 3.8 27B." Only the teaser was available, so the comparison basis is unknown. Co-Design Bench, which did not test Haiku, shows Qwen3.8-27B at 0.644 strict completion overall, close to frontier models. medium
• Which Qwen 3.8-27B file to download. The same author's teaser claims "The 17GB file coded as well as the 55GB original," across 24 quantized files on Hugging Face. The headline figures (17GB against 55GB, about 3.2x smaller) come from the teaser. The benchmark and sample size are not stated. medium
• Nandi for Indian-language complaint routing. A Towards AI post describes a free open-source model that sorts complaints in 13 Indian languages. Teaser only. Accuracy figures and license are not stated. medium
• Inside video generation on a real GPU. A Towards AI walkthrough of Sora and Wan2.1-style architectures covering 3D causal VAEs, "Spacetime Tubes," 3D RoPE and 4D attention. Teaser only. medium

Infrastructure

1 source, 1 min read

• The hidden cost of a database write. A DataDrivenInvestor explainer on write-ahead logging (WAL), fsync and durability. The basics still apply: durability guarantees are only as strong as the fsync behaviour underneath, as Binance's emergency "rigorous testing" of its withdrawal pipeline is a reminder that infrastructure assumptions only show their cost when they fail. Teaser only. medium

Science

1 source, 1 min read

• A "possible new planet" found with Claude. A Data Science Collective post says "a guy" found a possible new planet using Claude and insists it "is not a random clickbait story." Only the teaser was available. No paper, catalogue entry or confirmation by astronomers was cited in it, so treat it as an unconfirmed candidate at best. medium