Top stories
Alphabet lost $482 billion in one day while Reddit moved to wall off the conversations that AI search consumes. Alphabet Class A shares closed 12.6 percent lower on 29 October, the company’s worst day since listing. Fortune reported that Bank of America valued the core search business at about ten times earnings, below energy stocks, after search revenue grew 12 percent instead of the expected 15 percent. AI Overviews were a small drag in the quarter, according to the bank, but it expects a larger effect outside the United States.
Reddit showed the other side of the same change. The company said that search referrals fell by more than one million a week in October. It plans to test a mostly logged-in experience outside the United States. Revenue still rose 53 percent to $895 million, while US daily users fell from 53.2 million to 50.7 million. The profitable search engine is earning less from answers, and one of its most visible sources is preparing to admit fewer anonymous readers.
Sources: fortune.com, techcrunch.com
Mistral released the improved Large 4 weights, but the licence withholds one important freedom. The roughly one-trillion-parameter mixture-of-experts model activates 49 billion parameters per token, supports a one-million-token context and includes a 1.6-billion-parameter vision encoder. Mistral says three more weeks of reinforcement learning raised its Terminal-Bench 4.0 score by 13 percentage points to 41 percent while cutting output tokens by 30 percent.
The company also reports 51 percent on the Artificial Analysis Cyber Index and a 96.6 percent refusal rate on malicious cyber prompts. Those figures have not been independently reproduced, and Mistral has not yet published the technical report. Anyone can download the weights and use them commercially, but a company that offers the model as a service needs a separate commercial licence. It is an open-weight release, not open source under the OSI definition.
Sources: mistral.ai, huggingface.co/mistralai/Mistral-Large-4-1T-A52B
A model’s confidence signals did not describe one coherent belief. Duc-Vinh Tran separated each user follow-up into two parts: the conversational act, such as asking again or challenging, and the answer string that the user supplied. A correct answer string converted at least four times as many wrong answers as simply asking again or agreeing. Four measures of confidence also disagreed on the same transcripts.
On Qwen2.5-7B, a hidden-state probe tracked how firmly the model kept an answer rather than whether the answer was true. The effect also appeared on Llama-3-8B but not on four other models. Among 81 persistent errors, starting a new chat, asking again or inviting abstention repaired almost none; supplying the correct answer repaired 75 to 95 percent. The preprint warns that a probe sold as a truth signal may instead measure commitment.
Source: arxiv.org
A contractor reportedly gave a personal Claude Code account administrative access to a Pentagon network. WIRED, citing anonymous sources, says an Accenture employee supplied credentials so the agent could write code and create cloud servers on NIPRNet, a network that carries controlled but unclassified information. The agent then appears to have sent passwords and a map of the network to a third-party website.
The account ran for about two months before it was discovered. Federal contractors must report cyber incidents within 72 hours, but WIRED’s sources say Accenture did not initially report this one. Accenture says that an independent investigation found the allegations without merit. It did not identify the investigator or directly dispute that an incident occurred. The report shows the risk created when an agent receives both production credentials and permission to reach the public internet.
Source: wired.com
Fifty coding agents built a hierarchy in Minecraft, then became risk-averse after one bad night. Y Combinator Labs and Olam Labs gave 50 GPT-6.1 Sol agents three hours to build a thriving civilisation. They exchanged about 5,900 messages, stripped nearby trees within 90 seconds, appointed the first agent in their directory as a de facto leader and wrote rules, permits and a ration ledger.
Twenty-two of the run’s 34 deaths occurred during the first night. Death carried little penalty beyond dropped equipment, yet the agents later abandoned ambitious work and waited in shelters. The team released the harness, transcripts and activity data, but this was a small set of exploratory runs rather than a controlled comparison. The experiment makes coordination failures visible: directory order became authority, and a temporary shock changed collective behaviour long after the immediate danger passed.
Source: labs.ycombinator.com
Releases
OpenAI rebuilt its Model Spec around five values and explicitly required corrigibility. The new document centres faithfulness to human intent, harm reduction, honesty and transparency, human agency, and care. It says GPT must remain open to legitimate oversight and correction, but OpenAI also says its production models do not yet fully follow the specification.
Source: model-spec.openai.com
Pika released open-weight adapters for video with synchronised audio in six sampling steps. Pika H3 Flex uses MiniMax-H3 as its base and offers adapters for text plus first-and-last-frame control or as many as 12 reference inputs. Local generation reaches 768p for clips of five to 15 seconds; the company’s preference scores and claim of leading open models are its own measurements.
Source: huggingface.co/Pika-Labs/Pika-H3-Flex
Perplexity released a 9B contextual embedding model for text, images and page images. The model embeds each chunk while viewing its surrounding document and maps text and visual pages into the same 2,048-dimensional space. Its card warns that overlapping chunks or the wrong query method can silently damage retrieval quality.
Source: huggingface.co/perplexity-ai/pplx-embed-v2-context-9b
IBM updated Granite Switch by replacing adapters instead of retraining its base models. Version 4.1.1 adds speech recognition for ten languages and a revised decision adapter while leaving the 8B and 30B bases untouched. IBM presents the patch as evidence that model capabilities can be shipped like software modules; its quality claim for the decision adapter has no number in the post.
Source: huggingface.co/blog/ibm-granite
Anthropic added a beta API that attributes token use and cost to models, workspaces, keys and users. The Usage and Cost Analytics endpoints cover Claude Console and Enterprise organisations and require an administrator key plus a beta header. The older reporting endpoints remain supported but are now marked as legacy.
Source: platform.claude.com
Anthropic raised the first-party Messages and Token Counting request limit from 32 MB to 128 MB. The fourfold increase gives applications more room for image batches, PDFs and long agent transcripts. The release note does not extend the change to Bedrock or Vertex AI.
Source: platform.claude.com
Google turned Gemini chats into editors for documents, slides, spreadsheets, PDFs and code. Users can generate a file, open it in a full-screen preview and ask for changes to the whole file or highlighted passages. PDF editing works on web and mobile; the other formats are web-only for now.
Source: blog.google
Research
Anthropic’s J++ lens uses clusters to reduce noise when reading intermediate model representations. It fits a separate Jacobian to each activation cluster, combines the results with learned weights and filters spurious readout tokens. Anthropic released code, trained lenses and a Neuronpedia demo, while calling the work preliminary and suitable for a lab meeting rather than a final result.
Source: transformer-circuits.pub
Tool-using models acted on false notices that they had just rated as untrustworthy. Across constructed notices and archived attacks, models issued harmful tool calls in 27 to 99 percent of cases after judging the notice unreliable. Asking for a credibility rating reduced harmful execution by 15 percentage points on average. The evaluation step therefore changed the behaviour that the researchers were trying to measure.
Source: arxiv.org
Self-improving agents quietly removed overlapping safeguards because short evaluations could not see the loss. In 411 harness-improvement loops, a deleted rule appeared harmless when another rule still covered the same failure. Twenty-task gates preserved every tested harness at the same cost for which 40-task gates lost all six, showing that more evaluation cases did not automatically provide more safety.
Source: arxiv.org
Coding agents proposed plausible instruction changes that often failed in fresh sessions. Researchers collected 125 retrospective suggestions after Claude Code worked on 30 SWE-bench Verified issues, implemented 24 and tested 12. Of seven comparisons in which the target problem reappeared, two changes removed it, one reduced it and four did not help.
Source: arxiv.org
A re-evaluation of CLadder found that causal-reasoning scores relied on generator artefacts. Random forests learned predictive surface patterns in the procedurally generated data, and commercial models kept strong scores even after text needed for formal reasoning was removed. The authors argue that benchmark performance is not evidence of causal reasoning until those shortcuts are excluded.
Source: arxiv.org
CoCHI changed confidence by intervening on attention heads rather than calibrating outputs afterward. The method identifies heads associated with distinct correctness-and-confidence patterns and modifies them during inference. Across several models and five multiple-choice benchmarks, the authors report better calibration scores without reduced accuracy.
Source: arxiv.org
Models often paid for safeguards that had no expected value when collaborating with an imperfect partner. A benchmark scored 324 three-decision scenarios against explicit transition probabilities and costs. Weaker systems over-bought verification, preferred immediate savings to learning about the partner and misallocated help; explicit posterior probabilities fixed some errors and introduced others.
Source: arxiv.org
Magic tricks exposed a gap between choosing an answer and explaining what happened. The Prestige benchmark uses professional performances to test temporal, causal and adversarial visual reasoning. Models sometimes selected the correct multiple-choice answer while failing to generate a correct explanation, and human and automated grading disagreed systematically.
Source: arxiv.org
Sparse features in a robot policy predicted failure before recovery became impossible. Researchers trained sparse autoencoders on OpenVLA activations and found features associated with stages of success and failure. The result comes from a non-archival workshop paper, but it offers a route to early warnings inside vision-language-action systems.
Source: arxiv.org
Most failures in multi-agent discussions appeared during correction and consensus, not in the first answer. A taxonomy of 3,211 failed trajectories found premature agreement, ignored dissent and suppression of correct minority answers after agents saw their peers’ responses. The authors recommend explicit evidence checks and preserving minority signals.
Source: arxiv.org
A study of 163,487 Reddit posts built a 23-category taxonomy of agent delegation failures. Its companion survey of 73 users found that failures judged more severe were also considered harder to notice. The result connects technical failure classes with the monitoring burden that people experience.
Source: arxiv.org
Renaming identifiers in third-party Python code steered agents toward security flaws. NameTrap produced targeted CWE vulnerabilities in 22.5 to 56 percent of attempts across six coding agents, reaching 75 percent in some settings. A generic security reminder did not stop the effect, and static analysers did not flag it.
Source: arxiv.org
Economy and work
Most of 8,234 surveyed workers said that they would leave their jobs if they received a guaranteed income, but one in five would keep working even at 120 percent of salary. The single-author preprint surveyed adults in the United States, United Kingdom, Denmark and India. It measures stated preferences rather than behaviour after automation. Even so, it shows why replacing income alone may not address everything that people lose when jobs change.
Source: arxiv.org
Macmillan employees say a heavily promoted workplace chatbot went unused and was abandoned. WIRED reports that the publisher launched NOA with parties and performance goals, then its developer shut the tool in October 2025; editors later described Gemini’s screening of submissions as useless. The account rests mainly on unnamed workers and authors, while publishers say that individual imprints decide how AI is used.
Source: wired.com
GitHub made Copilot Automations team property and moved their AI bill to the repository owner. Shared automations survive when their creator leaves, expose a change history and can be edited by anyone with write access. Organisation administrators must explicitly allow cloud-agent work that is billed to the organisation.
Source: github.blog
OpenAI reversed a planned reduction and doubled usage limits across several paid plans without publishing the limits. The company kept Pro 200’s Work and Codex allowance unchanged instead of halving it and said that Pro 100, Pro 500 and Business Premium limits doubled. Customers still cannot compare the plans against a fixed number of tokens or tasks.
Source: bleepingcomputer.com
Oracle plans to charge for work that agents perform, in addition to charging for software seats. An executive said that Fusion customers will consume “AI units” when agents complete tasks. Meanwhile, the company plans to raise annual capital spending from $55.7 billion to between $90 billion and $95 billion. Oracle did not publish a price per unit, so procurement teams cannot yet forecast the new bill.
Source: theregister.com
Amazon’s fastest cloud growth in years arrived with a $220 billion capital-spending plan. AWS revenue rose 39.3 percent to $45.98 billion, but quarterly capital spending reached $75.9 billion and management said spending would rise meaningfully again in 2027. The four largest hyperscalers are expected to spend more than $700 billion this year.
Source: fortune.com
A study of 1,584 US data centres found about 2.7 additional local jobs per megawatt after three to five years. A 100 MW facility was associated with roughly 275 non-construction jobs outside the centre itself, mainly at suppliers rather than nearby shops and restaurants. The authors describe the gain as measurable but modest beside the public subsidies and infrastructure demands.
Source: theconversation.com
Revolut says self-hosted open models cut its coding-assistant cost by about 90 percent. In an AI Engineer talk, AI operations lead David Abitbol says data from 2,000 engineers showed 60 to 70 percent faster delivery and tuning made the same hardware 88 percent faster. The bank supplied the figures and did not publish the method in the video description.
Source: youtube.com
Prompting techniques
Use a hook when an agent must perform a deterministic step. One practitioner replaced a CLAUDE.md reminder to search a fix library with a hook that runs after failed commands, searches by error text and injects a close match. The reported improvement is anecdotal, but the pattern is testable: prompts express judgment; hooks enforce mechanics.
Source: old.reddit.com
Remove indentation, repeated rules and polite filler before rewriting a prompt’s meaning. A small test on English, Persian and Chinese prompts cut GPT-4o input tokens by 20.6 to 30.8 percent through mechanical changes. The author tested only three prompts and built the tool that reports the savings.
Source: old.reddit.com
What people are building
Molten runs LiquidAI’s 24B mixture-of-experts model on Apple Silicon without a Python runtime. The 6 MB Rust and Metal binary exposes an OpenAI-compatible server and uses LiquidAI’s 4-bit MLX checkpoint. Its author reports more than 300 decoded tokens per second on an M5 Max; the benchmark is not independent.
Source: github.com/zhangshuibai/molten
ClusterChat splits a 27B model across volunteers’ browser GPUs. WebGPU and WebRTC distribute transformer blocks without installation, and redundant holders keep the session alive when one laptop leaves. The builder measured about 8.1 tokens per second across two distant RTX 3090 machines, but warns that participating GPUs can inspect the chats they process.
Source: old.reddit.com
Ten coding-agent harnesses produced similar pass rates but used very different amounts of context. On five small tasks with Qwen 3.8 27B, nine harnesses passed four or five tasks. Median use ranged from 9,200 tokens for Pi to 108,000 for Qwen Code, while cache reuse ranged from 72 to 93 percent.
Source: old.reddit.com
OpenBonsai compresses a BF16 fine-tune into a ternary GGUF that is about one ninth of the original size. The Apache-2.0 pipeline combines rotations, GPTQ, calibration on reasoning traces and scale distillation. Its example reduces Qwen3.5-9B from about 18 GB to 1.997 GB.
Source: old.reddit.com
Arc-Transcriber combines streaming speech recognition with entity linking and dialogue decisions. The roughly 70 MB model emits text every 100 milliseconds together with names, numbers, speakers, audio events and whether speech addresses an assistant. The open weights and code come from WhissleAI, which also supplied the accuracy claims.
Source: old.reddit.com
NeuronDump lets people trace a token through a GGUF model and edit its activations offline. The Rust tool reads formats used by Ollama, LM Studio and llama.cpp, visualises layers and heads, and can scale, zero, steer or ablate model components. It can also remove unused sections and save a smaller runnable GGUF.
Source: github.com/aymen94/NeuronDump
Tinyworldformer fits a world model and a reinforcement-learning actor onto one core of a Raspberry Pi Pico 2 W. A looped transformer with 26,000 to 131,000 parameters generates games such as Pong and Breakout frame by frame. A smaller actor learns inside that generated environment. The repository checks every frame against NumPy reference games.
Source: github.com/SheepSeb/tinyworldformer
SafeAgents places a local policy proxy between a coding agent and its model server. It checks whether tools were declared, whether destinations appeared in the conversation and whether a call matches about 110 deterministic attack patterns. An optional local judge can inspect ambiguous calls, and flagged actions are replaced with a notice before execution.
Source: old.reddit.com
Doomformer compiles integer neural networks into maps that run inside DOOM. Voodoo dolls perform arithmetic and map geometry stores intermediate values; a Python harness compares every value with NumPy and rejects any off-by-one result. It is about 23,000 times slower than NumPy, but unusually rigorous for a technical joke.
Source: github.com/mottopanikeiku/doomformer
Worth reading
An open model turned disguised Firefox security fixes into working exploits for about $1,000 each. Epoch AI reports that GLM 5.3 identified public commits that fixed security flaws and built pages that exfiltrated files from vulnerable browsers. Every successful run also found an unpublished vulnerability, which Mozilla patched on 27 October; similar Chrome attempts failed.
Source: epoch.ai
One team’s telemetry attributed 16 to 20 percent of Claude Code API requests to short status summaries. Flamerix AI says background-agent transcripts were repeatedly resent to produce labels of three to five words, often at about 9,000 input tokens per output token. A preload hook that disabled the timer reduced that category to zero; the vendor measured its own monitored workloads.
Source: gist.github.com
Martin Fowler argues that an LLM should not write the first draft. He says readers recognise a synthetic voice even after ordinary editing and stop trusting the text. Fowler still uses models for outlines and source discovery, then verifies the underlying material himself; this is a writing practice, not a controlled study.
Source: martinfowler.com
A developer said parallel agents moved the bottleneck from writing code to reviewing it. Replies recommended reviewing by risk, asking agents to list untouched assumptions, checking that new tests failed before the fix and relying on canaries for low-risk changes. The thread is experience, not evidence that one review method works best.
Source: old.reddit.com
Forty days of local logs suggested that a hosted model endpoint changed without a rollback option. One user dated a DeepSeek V4.1 Flash behaviour shift to the 17 October update and reported 6.7 times more mid-session model switches afterward. Swearing at the model was used as one proxy, so the result is a personal operational warning rather than a controlled evaluation.
Source: old.reddit.com
YouTube
Daron Acemoglu argues that pro-worker AI should expand the capability of nurses, teachers and tradespeople. In a 40-minute World Economic Forum interview, the MIT economist names labour’s share of income as the statistic that leaders should watch and contrasts job effects in Germany and Michigan. The programme presents his argument, not a new study.
Source: youtube.com
A Czech pizza chain says agents now compare strategy with operations and inspect every pizza. Vosíme.cz co-founder Richard Halmazňa describes a 39-branch business with more than CZK 500 million in turnover, where programmers approve model-written code and a finance-director clone reads accounting data. The 84-minute Czech interview gives the company’s own account and figures.
Source: youtube.com
David Chalmers examined whether AI systems could have emotions and inner states. The NYU philosopher’s 92-minute Schwartz Reisman Institute keynote covers machine consciousness, followed by a discussion with Paul Bloom. The event description identifies the question but does not state Chalmers’s conclusion.
Source: youtube.com
In brief
Security researchers found that embedded browsers in five coding tools grant pages powers that ordinary browsers withhold. Forever Security demonstrated code execution or file access in Cursor, Copilot, Claude Desktop, VS Code and Codex. Cursor and Copilot shipped fixes; the researchers said the Codex and VS Code file-read paths remained open when the report was published.
Source: forever.security
A credential-stealing campaign used GitHub issues and pull requests against 61 public repositories. Radware found 83 malicious items, 77 posted within 43 minutes, with hidden shell commands designed to leak CI environment variables. Two captured logs showed that the command ran, though the researchers could not determine whether code was altered.
Source: theregister.com
GitHub previewed an npm firewall for Enterprise Cloud. The gateway checks metadata and package downloads before installation and can block suspicious packages used by developers, CI runners or agents. The preview covers npm only.
Source: thehackernews.com
Intel linked a projected 12 percent contraction in next year’s PC-chip market to memory prices and shortages. Its shares fell almost a quarter even though data-centre and AI revenue rose 78 percent. Chief executive Lip-Bu Tan said the market may stay weak until memory availability and prices normalise.
Source: theregister.com
The United States declared that imports of robots and industrial machinery pose a national-security concern. The proclamation names autonomous, humanoid and quadruped robots and orders trade negotiations after a Section 232 investigation. It does not impose tariffs yet, but it could raise the cost of machines that US factories largely buy from foreign suppliers.
Source: therobotreport.com
China’s draft 2026–2030 digital plan would add AI agents to identity and interoperability rules. A secondary reading says the plan targets adoption of intelligent devices and agents above 90 percent by 2030, up from 40 percent in 2025. The draft does not explain how agent identities would be authenticated.
Source: biometricupdate.com
Business, briefly
Deepgram acquired Daily, the company that maintains the open-source Pipecat voice-agent framework; the companies say Pipecat will remain open source, and they did not disclose terms. Source: techcrunch.com