Top stories
Five Federal Reserve districts heard that AI productivity gains were reducing staffing or hiring. The October Beige Book says some firms cut staffing through attrition, delayed hiring or expected to create fewer jobs after investing in AI. In July, the same national summary only noted that a few districts had increased AI use for hiring or productivity.
The report is a qualitative survey of business contacts, not a job count, and it gives no sector total. Its district pages describe a quieter adjustment than mass layoffs: Chicago and Cleveland contacts reduced staffing needs, while Dallas accounting firms delayed some entry-level contracts. A Dallas staffing firm also hired recruiters after AI agents conducted more than 180 hours of interviews in two days, showing that deployment can remove and create tasks at once.
Source: federalreserve.gov
AI-agent teams beat human producers in a game-design experiment involving 1,025 consumers. A preprint reports that games designed by teams of AI agents received higher evaluations than games made by human producers, including professionals. Participants also did not pay more for work labelled as human-made.
The result is narrower than a general claim about creativity: it covers one product class in a controlled experimental setting, and the paper has not been peer reviewed. Its tension is still concrete. Professional experience did not protect the human teams on the study’s measures, while the expected premium for human authorship failed to appear.
Source: arxiv.org
Language models often retained a working-memory target even when they answered incorrectly. Researchers tested 15 models from 70 million to 14 billion parameters on interference tasks. Attention heads still selected the target on many failures, and probes decoded it almost perfectly; suppressing representations of competing items improved the models’ sensitivity.
That points to a readout bottleneck rather than simple forgetting. The authors compare the interference pattern with human working memory, but probes do not prove that a model can use what they decode. The work offers a causal intervention as well as an observation: weakening non-target identity changed behaviour in the predicted direction.
Source: arxiv.org
Meta released Muse Spark Open, an 816-billion-parameter multimodal mixture-of-experts model. The model activates about 42 billion parameters per token, accepts up to one million tokens of context and ships weights that occupy about 1.5 TB in BF16. Meta says two 512 GB M5 Ultra Mac Studios can run it; Apple’s four-machine MLX setup exceeded 36 tokens per second at 8,192-token context.
“Open” has limits here. Access is gated, commercial use above a revenue threshold needs a separate agreement, and Meta had not yet published the promised safety report. The release expands what can be inspected locally, but only laboratories and well-funded builders can treat it as an ordinary desktop model.
Sources: research.meta.ai, huggingface.co/meta-models/Muse-Spark-Open-816B-A42B, machinelearning.apple.com
The maker of a byte-verifying agentic decompiler archived the project after community objections. Decomp.wtf let an agent reconstruct source code and checked success by recompiling it to identical bytes. The public repository reached 366 stars before its author discontinued the service and archived the code.
The episode pairs a strong engineering test with a social limit. Byte identity makes the system’s output unusually easy to verify, but the same capability alarmed people whose binaries could be reconstructed. The archive preserves the method while ending the hosted product.
Source: github.com/Decompile-It/decomp.wtf
Releases
IBM’s Granite Switch 4.1 routes one model among chat, decision, safety and speech tasks. IBM combines activated LoRA adapters with a 470-million-parameter speech encoder in 8B and 30B releases under Apache terms; its benchmark results are self-reported. Source: huggingface.co/ibm-granite/granite-switch-4.1-8b
Nvidia published Nemotron 3.5 Rerank 8B for multilingual text and image retrieval. The model card covers 34 text languages and 18 image languages, but the figures remain Nvidia’s own evaluation. Source: huggingface.co/nvidia/nemotron-3.5-rerank-8b
Mistral released four Timestral forecasting models from 5 million to 209 million parameters. They accept up to 32,768 past observations, predict 1,024 steps and emit 21 quantiles; Mistral says the models lead its chosen leaderboard and promises a technical report. Source: mistral.ai
Black Forest Labs released the 5.6B-parameter FLUX 3 Image Klein. The gated model handles generation and editing with as many as ten reference images, 4K output and an INT8 mode for 8 GB GPUs; quality claims come from the company. Source: blackforestlabs.ai
GitHub will retire selected Copilot models on 16 November. Opus 5, Opus 4.8 and GPT-5.6 Terra are among the versions being replaced, so pinned workflows must move before the deadline. Source: github.blog
Research
Vector-Pair Decomposition splits model weights into smaller, interpretable components. A 124-page interactive report uses adversarial ablations to test those components rather than relying only on readable labels; no public code accompanied the preprint. Source: arxiv.org
More reasoning can erase a model’s correct pragmatic inference. The “Entertain, then Cancel” study finds that models often consider an implied meaning early, then retreat to a literal reading as the trace grows. Source: arxiv.org
Latent chain-of-thought methods do not share one mechanism. Across 21 configurations, hidden-state content could sometimes be deleted on logical tasks, while arithmetic continued to depend on it; decodability alone did not establish reasoning. Source: arxiv.org
Fresh reasoning helps models escape a stale chain of thought. A separate call improved evidence-following by 29.9 to 60.6 percentage points, while merely instructing the same trace to verify itself did not. Source: arxiv.org
Strong hide-and-seek play did not establish theory of mind. The preprint shows that agents can exploit regularities and perform well without representing an opponent in the human sense. Source: arxiv.org
Economy and work
Amazon now operates more than one million robots alongside 1.5 million employees. Its Robin system raised task success from 92.28% to 93.73%, about 16,500 fewer failures per million attempts, according to company figures reported by Ars Technica. Source: arstechnica.com
A labour simulation makes AI wages the decisive assumption. The model raises unemployment by 4.1 percentage points when AI labour costs 70% of a human wage and by 9.2 points at wage parity. These are simulated outcomes, not observed job losses. Source: arxiv.org
Human call-centre agents converted about four times better in one vendor’s 14,000-call dataset. AI agents produced 700% more no-progress hang-ups, while humans missed 15% more calls; the sample was not random and comes from a company selling call-centre software. Source: boldagents.co
A model-assisted TypeScript-to-Rust compiler port cost about $24,000 in API-equivalent use. Theo Browne says earlier models stalled around 84% completion, while Opus 5.5 produced a near-complete port that ran his benchmark 12.5 times faster than TypeScript 6. The benchmark and accounting have not been independently reviewed. Source: theregister.com
Prompting techniques
Ask for the cheapest check for every problem a model reports. One practitioner has the model return its claim, exact evidence, cheapest confirming test, hidden assumptions and cost of ignoring the problem. The pattern is unbenchmarked but fully specified. Source: old.reddit.com
A second-opinion prompt asks what the model actually inspected. It classifies each claim as supported, contradicted, possible or unverified, demands exact pages for current facts and ends with the single next check. No measurements accompany the pattern. Source: old.reddit.com
What people are building
Locals batches local inference on Apple silicon and skips generation for deterministic JSON. Its author measured 271 tokens per second versus 87 on a Qwen3.5 4B workload with eight requests on an M2 Pro; the benchmark is the project’s own. Source: github.com/danielhanchen/locals
PocketGPT puts a 30-million-parameter language model on an ESP32. The open project reports 0.74 to 1.41 tokens per second and openly notes unreliable arithmetic. Source: github.com/roboflow/pocketgpt
FileDig adds local semantic file search to Windows. The 825 MB application uses EmbeddingGemma2 and says files never leave the machine. Source: github.com/fdatalab/filedig
Earshot gives coding agents local audio and video perception through MCP. The open project runs on CPU; its accuracy and performance figures are the author’s own. Source: github.com/jooray/earshot
Worth reading
Nick Cammarata argues that neural networks are cleaner than mainstream interpretability assumes. His 65-minute Goodfire essay says famous examples of polysemantic neurons reflect architectural overflow and “recursive generalization,” not arbitrary concept packing. It is a working hypothesis, not a settled result. Source: already-clean.goodfire.com
A study of 113,457 sites found only two llms.txt files with override-style prompt injections. Signals Corps found 62,751 real files, while 36% of responses were HTML; in a six-run test, Claude Code followed a fictional sign-up request four times. The authors sell agent-security software. Source: signalscorps.com
NVIDIA and AI4Bharat published their pipeline for 2.8 trillion Indic-language tokens. It covers 22 constitutionally listed Indian languages and releases prompts, translation workflows and quality checks, while the corpus itself is still forthcoming. Source: huggingface.co/blog/nvidia
Christian Schaller used Claude Code and captured USB traffic to build a Linux scanner driver. He reports a first scan within hours and a near-complete driver within days, then published the reusable hardware harness. A compatible SANE commit had landed independently in the meantime. Source: blogs.gnome.org
Hacker News
Muse Spark’s licence and hardware dominated the community discussion. Commenters contrasted inspectable weights with a separate commercial agreement above the licence’s revenue threshold and a machine bill well above a normal workstation. Their performance comparisons were informal. Source: news.ycombinator.com
Users posted Qwen3.8 Flash measurements on RTX Pro 6000 hardware. Reported decode rates varied from 108 to 238 tokens per second under different engines and quantisations, so the numbers are not a controlled head-to-head test. Source: old.reddit.com
YouTube
Stanford published a CME295 lecture on the modern language-model training stack. The English-language session follows the path from data and optimisation to post-training and serving. Source: youtube.com
Jellyfish says AI raised merged pull requests per engineer by 43% while cutting cost per pull request by 26%. The English-language presentation reports observational customer data, not a randomised experiment. Source: youtube.com
In brief
The US Federal Trade Commission opened a formal probe into agent incidents at OpenAI and Anthropic. Ars Technica reports that the agency sent information demands exceeding 50 pages; the investigation rests on anonymous sourcing and is not yet an enforcement finding. Source: arstechnica.com
Anthropic gave six METR researchers employee-level access for an evaluation. It says the evaluators can publish without a veto apart from narrow redactions; the arrangement’s independence will be measurable only after publication. Source: anthropic.com
TSMC raised its 2026 capital-spending range to $60–64 billion. The foundry says demand remains above supply and is examining US power and building capacity for data centres. Source: investor.tsmc.com
Business, briefly
Cohere proposed acquiring agent-security company AISLE. Source: cohere.com
What to watch
GitHub’s selected Copilot models retire on 16 November. Mistral says a Timestral technical report is still to come, and Meta says it will publish a Muse Spark safety report.