Top stories
AI use at work doubled across the euro area while hiring plans weakened, especially for graduates. ECB executive-board member Philip Lane said 52% of workers now use AI at work, up from 26% in 2024. Separate ECB business data show 64% of firms use it experimentally or moderately, but only 7% use it significantly.
The central bank sees no large aggregate employment or productivity effect yet. Time savings cluster in coding and other specialised tasks, while muted hiring falls first on new entrants. Lane framed the outcome as a race between labour costs and token prices and warned that AI could weaken workers’ bargaining power and labour’s share of income.
Source: ecb.europa.eu
One persistent adversarial agent raised false private beliefs among benign agents from 0.6% to 44.8%. In factual discussions among ten agents, the adversary repeatedly asserted a plausible wrong answer. Three adversaries pushed the final false-answer rate to 60.5% across six models and three question-answer datasets.
The internal result is stranger than ordinary public conformity. With three adversaries, benign agents privately held the false belief in 51.6% of exposed turns but publicly endorsed it in only 19.7%. Agents that retained the truth often failed to defend it, and baseline factual knowledge did not predict resistance. The result directly challenges the assumption that a group of agents naturally checks itself.
Source: arxiv.org
OpenCode says attackers may have taken provider keys stored in its hosted Console. The breach affected a legacy system being decommissioned and 1.4% of Console users, according to the company. The affected product is the hosted routing service, not the local coding agent; users who received a notice should revoke stored OpenAI, Anthropic and Google keys and inspect provider logs.
OpenCode called the incident “large scale distillation,” but has published no evidence for that motive, no cause and no exposure window. One Reddit user reports that a stored Anthropic key drained $158 overnight. The breach shows the extra trust created when a gateway holds credentials that can spend directly at another provider.
Source: news.ycombinator.com
DeepSeek updated V4.1 Flash with higher self-reported agentic-coding scores at unchanged API prices. The lab reports Terminal-Bench 4.0 rising from 31.2 to 42.3 and ProgramBench from 20.3 to 29.1, with shorter outputs and less task time. The MIT-licensed checkpoint accepts one million tokens; the API remains $0.15 per million uncached input tokens and $0.60 per million output tokens off peak.
An independent early coding test passed all 30 runs at about $0.045 per task, but used about 550,000 tokens per task against roughly 30,000 for Opus 5.5 on low effort. That single test complicates the lab’s efficiency claim: the model is cheap per token and capable, but may use far more tokens.
Source: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
Xolotl runs a 552-billion-parameter DeepSeek mixture-of-experts model on one RTX 4090, system RAM and NVMe. The Strata-based engine loads only the experts needed for each token across 24 GB VRAM, 90 GB of DDR4 and disk. Its author reports 27 tokens per second, 180–320 prompt tokens per second and 512K context on a Ryzen 7 5700X machine.
The same setup passed the author’s small tests—21 HumanEval tasks, ten tool calls and three agentic bug-fix loops—but these are self-measured results on an older quantised checkpoint. It is still a striking access result: a model normally associated with server-class memory runs interactively on consumer hardware by trading memory hierarchy and storage bandwidth for latency.
Source: github.com/Indras-Mirror/Xolotl
Releases
DeepSeek opened its in-house Search and Fetch APIs. Each endpoint will cost $0.75 per 1,000 successful calls from 1 November; calls made as server-side tools through DeepSeek Harness, Claude Code or Codex remain free until then. Source: api-docs.deepseek.com
Research
In-context learning develops as several capabilities, not one. Across 17 models, checkpoint analysis separated prompt control, answer preference, construction and transfer; models attended to task-relevant tokens before their accuracy rose above chance. Source: arxiv.org
Whether a code behaviour looks localised depends on the probe and component budget. Across ten models, only one of 15 settings beat random controls at budget 12, versus six at budget 40, reversing some earlier claims that interventions uniformly fail. Source: arxiv.org
Reasoning density predicts correctness differently after supervised and reinforcement training. Forcing the measured density cut accuracy from 60.0% to 8.3% on a 60-question GSM8K sample, while a hint plus a cache intervention added 13.3 points. Source: arxiv.org
A 65-property semantic representation explains both shared and unique brain variance. In seven people listening to stories, concrete properties aligned more strongly with LLaMA2-7B, while interpretable features retained brain-predictive information the model missed. Source: arxiv.org
Sparse-autoencoder features are more additive than random coordinates but still interact. Adding pairwise effects raised steering rank fidelity from 0.587 to 0.779, warning against explanations that treat learned features independently. Source: arxiv.org
Automated research agents improved when they periodically forgot their history. Restarting from the best candidate found so far gained up to 15% over communicating parallel agents and 39% over a single agent across four tasks. Source: arxiv.org
Economy and work
Amazon’s Atlas system scores warehouses from 100 to 900 for unionisation risk. Internal documents cite hundreds of inputs, including Reddit posts and survey answers, and place 16 sites in the highest-risk tier. Amazon says Atlas works only at site level and does not monitor individual views or union membership. Source: opendemocracy.net
Gartner estimates data-centre owners need $8.5 trillion in cumulative revenue through 2030 but will receive at most $3.6 trillion from AI. Bank of America separately says business AI spending has begun to dip as lower token prices outweigh higher volume. These are analyst estimates reported by Fortune, not audited accounts. Source: fortune.com
Banks, retailers and airlines are blocking consumer agents or charging them for access. JPMorgan, Schwab, Amazon and eBay block agents; Delta prevents automated booking and Yelp admits only agents that license its data. Meta is negotiating access retailer by retailer. Source: theverge.com
China’s integrated-circuit export value rose 136% while volume rose only 1%. September exports reached $45 billion for 32.7 billion units, according to customs data reported by SCMP, showing how the memory shortage moved revenue through price rather than shipments. Source: scmp.com
HUMXN offers free home repairs in return for filming skilled work to train robots. Technicians wearing capture sensors record inspection, tool choice and completion in Minneapolis, Chicago and Miami. The company gives no worker-payment terms, dataset size or robot-deployment figures. Source: tomshardware.com
Prompting techniques
Nine system-prompt rules raised one model’s Terminal-Bench score from 52.8% to 59.6%. The rules tell the agent to finish one approach, challenge false premises, treat checked answers as settled and seek outside evidence rather than repeatedly doubting itself. It is one 89-task run on GLM-5.3-Flash. Source: old.reddit.com
What people are building
TensorFold ran DeepSeek V4.1 Flash at 56.8 tokens per second across two DGX Spark systems. The author’s EXL3 recipe more than doubled text throughput against the vLLM kit on the same hardware and added persistent prompt caching; results are self-measured. Source: github.com/Elevatormusic/DeepSeek-V4.1-Flash-1017-EXL3-TensorFold-2x-DGX-Spark
Llmspeed simulates how a local model will feel on a selected GPU or Mac. Its estimates mix community measurements with projections; owners in the Hacker News thread found large errors and missing harness details, so it is a rough planning aid. Source: llmspeed.dev
A Ternary Bonsai agent runs entirely in a browser tab with WebGPU. The static Hugging Face Space offers 2B, 4B, 9B and 27B 1.58-bit models without a server and makes no benchmark claim. Source: huggingface.co/spaces/webml-community
Scout turns a plain-language task into clicks, typing and scrolling in a real browser. The Apache-2.0 agent accepts the user’s model keys and runs locally or on Outcrawl’s paid browser service; the README gives no reliability benchmark. Source: github.com/outcrawlai/scout
E2E records a verified agent step and replays it later without another model call. Ordinary assertions validate the first run; subsequent web or mobile tests reuse actions until the interface changes, reducing recurring inference cost. Source: github.com/tester-army/e2e
Reason Network lets teammates’ coding agents exchange messages peer to peer. The Rust daemon exposes team sessions, inboxes and status through MCP with no central server and supports Claude Code, Codex, Cursor and other harnesses. Source: github.com/reason-machines/reason-network
Worth reading
Nathan Lambert argues Meta and Reflection halved the Western open-model gap. His index still places more than twenty Chinese open models ahead without those releases, while he says similarly capable Western models attract more adoption. This is an evidence-backed argument, not a benchmark paper. Source: interconnects.ai
Ben Recht calls OpenAI’s faster Fourier transform a win on an unrealistic computational model. He argues its hidden constants and machine assumptions make the result closer to exploiting the benchmark than accelerating a real FFT; this is his technical critique, not a formal refutation. Source: argmin.net
Sam Newman describes disaster-recovery patterns for software that depends on AI vendors. The book excerpt covers backup-and-restore, pilot-light, warm-failover and active-active designs under GPU scarcity and unpredictable provider availability. Source: newsletter.pragmaticengineer.com
Hacker News
Amazon’s Atlas system prompted debate about AI worker surveillance. Commenters with workforce-analytics experience said similar systems are often sold as attrition or engagement tools even when their inputs reveal organising risk. Their accounts are anecdotal. Source: news.ycombinator.com
Practitioners debated who verifies agent-written code. One commenter reported breaking bugs up about 40% year on year while features and revenue grew faster; others argued the apparent productivity shifts quality-assurance cost onto users. Source: news.ycombinator.com
An OpenCode Console user says a leaked stored key drained $158 overnight. The poster reports Anthropic initially refused a refund; claims that affected users will be reimbursed are secondhand. The practical response is key rotation, low spend limits and log review. Source: old.reddit.com
Strata users found defaults that traded quality for speed. Builds before 0.1.45 could run greedily when clients omitted sampling, while route-tail skipping raised one user’s speed from about 72–81 to 93 tokens per second but doubled a published divergence measure. Source: old.reddit.com
YouTube
Featherless AI presents four ways to cut inference costs. Eugene Cheah recommends measuring spend, right-sizing models, buying GPU capacity deliberately and checking whether token volume represents useful work; the figures are his company’s. Source: youtube.com
Arize builds an evaluation loop for a financial agent. The workshop combines traces, code checks, model judges and CI gates after an example agent confidently reports that “116% of analysts agree.” Source: youtube.com
Monzo engineers argue spreadsheet calculations belong in code, not the model. Their live financial-agent workshop centres evidence, consistency, accountability and verification rather than response speed. Source: youtube.com
The New York Times profiles workers filming trades to train robots. Workers wear cameras while doing physical jobs, producing the data intended to automate parts of those same jobs. Source: youtube.com
In brief
Anthropic’s draft US bill would permit a model block only for imminent risk above $1 billion or more than 50 people killed or seriously injured. The reported text covers biological weapons, control evasion and cyberattacks while pre-empting some stronger state rules. Source: motherjones.com
Meta removed the word “train” from new commercial crawling terms after criticism. The company says its crawler behaviour has not changed; publishers note that training is no longer listed but is not expressly prohibited. Source: theverge.com
Nvidia denied discontinuing the RTX 5090 as supply tightened. Tom’s Hardware found the card absent from several major stores and offered around $8,500 on Amazon, but Nvidia says production continues. Source: tomshardware.com
Apple chief executive John Ternus called the memory-chip shortage “absolutely a crisis.” The shortage affecting AI servers and GPUs is now constraining consumer devices; reporting says Apple has tested Chinese memory suppliers amid US political pressure. Source: thenextweb.com
What to watch
DeepSeek begins charging $0.75 per 1,000 Search or Fetch calls on 1 November. Meta’s revised commercial terms take effect on 30 October.