読書

Belgian road-safety group Johanna.be reported that FSD exceeded the limit in most 30 km/h zones tested around Brussels, averaging 44 km/h, and attempted to overtake cyclists where passing is prohibited.

The test ran over three days in July, covering roughly 400 kilometers. The report also said the system's screen frequently displayed incorrect speed limits, though it acknowledged FSD behaved cautiously around pedestrians and cyclists, often yielding proactively.

The tester is an advocacy group, not a regulator, and the sample was limited; Tesla released its own study this month saying FSD does not speed excessively. A Flemish transport ministry spokesperson noted the driver remains fully responsible. An EU-wide vote could take place as early as next month.

Three of the 5,000 leaked FBI employee records explicitly reference the Remote Operations Unit (ROU), the FBI's secretive hacking team.

404 Media verified parts of the data by cross-referencing OSINT tools. The records include home addresses, multiple phone numbers, and in some cases spouses' names and their phone numbers. Reuters reported some job titles relate to investigating China or Russia.

The FBI says the point of breach is still undetermined — whether a third party or the FBI's own enterprise — and it is investigating. Only parts of the list have been verified, so readers should keep that margin.

Amazon's chief sustainability officer Kara Hurst said at an Axios event that the company does not know how it will achieve its goal of net-zero carbon emissions by 2040.

Per TechCrunch, Hurst said "we are still striving towards that target" but she is "not going to sit here and say, 'We know all the ways that we're going to do this.'" She confirmed The Climate Pledge commitment has not changed.

Amazon says it has been one of the world's largest purchasers of clean energy for several years, with 42 gigawatts in its portfolio, yet it was earlier this year reported to be planning a 7.65 gigawatt natural gas plant for an AI data center campus. The gap between an unchanged pledge and an unknown path is the signal to track.

A Gallup survey found that 68% of Americans who use AI daily are worried about the technology, against 32% among all respondents.

The study covers 37 countries with roughly 1,000 respondents each, polled between April and July. The US is the most anxious: 74% of Americans reported worry and only 36% expect AI to mostly help the country, while high-adoption countries like Singapore and China skew far more optimistic.

Note the basis: the study was commissioned by Microsoft, and what is available is TechCrunch's reporting on the results. The takeaway for readers is that usage frequency is not a proxy for public acceptance, and AI deployment in hiring, healthcare or government may face more resistance than adoption numbers suggest.

Gita Gopinath says China's large surpluses reflect things going wrong inside China, not strength.

Gopinath was first deputy managing director and chief economist at the IMF and has returned to Harvard. Speaking on Tyler Cowen's Conversations with Tyler, with a transcript published, she said the imbalance itself is not what matters — jobs, consumption and purchasing power are.

She pointed to weak Chinese consumption, misallocated resources, a property market unfixed for five years, and anti-involution price coordination in sectors like EVs. She also noted that rising imbalances have historically triggered trade wars, as in the 1980s Plaza Accord era. This is her judgment, not new data or a policy change.

Six named researchers published an essay on LessWrong on September 23 arguing that a shift to latent reasoning architectures would sharply undermine chain-of-thought (CoT) as an oversight tool.

The essay names three architecture families: COCONUT, which would replace CoT entirely; full-bandwidth transformers, which would add a latent channel alongside CoT; and looped transformers, which deepen computation between text bottlenecks. The authors say such changes 'could very plausibly become adopted in the near future,' as vendors may trade monitorability for competitive advantage.

The authors distinguish two oversight mechanisms: models needing CoT to complete hard tasks, and models tending to verbalize plans even when unnecessary. They judge the necessity argument robust for several years absent architectural change, while the propensity argument is fragile under adversarial pressure; latent architectures would weaken both. Note this is an analysis piece, not a new experimental result — the authors themselves call the necessity case 'far from guaranteed,' and no vendor has announced adoption of such architectures.

Holding AI providers liable may backfire: compensation weakens defense and raises attacker profits, and under monopoly with sufficiently many productive users, zero liability is uniquely optimal for a range of parameters.

Policy discussion has assumed provider liability improves safety, but this theory shows liability can improve welfare while increasing harm: a higher common price reduces effort on both sides without changing attack success, while compensation weakens defense.

This is a theoretical result from an SSRN preprint by Joshua Gans, economist at the University of Toronto, not empirically tested, per Tyler Cowen's blog summary.

The result is purely theoretical, with no empirical support and no independent replication; the preprint was posted to SSRN around 2026-09-23.

Readers building agent memory systems now have an evaluation option that puts accuracy, total cost and latency on equal footing: DolphinBench requires all three, with none optional.

Before this, evaluations typically looked only at accuracy, leaving cost and latency unreported and making the real price of a memory method hard to judge.

The benchmark has three knowledge-work personas with roughly 500k tokens of user messages each, and 200 tasks per persona that depend on that history; the authors verified each task by requiring an agent to succeed with the relevant history and fail without it. This is first-party design and verification.

There is no independent reproduction or third-party adoption yet; the preprint material is dated 2026-09-21 and can be tracked as an evaluation option.

Pretraining on purely machine-translated text reportedly reaches state-of-the-art results on French life-sciences downstream tasks, accompanied by a released 36.4GB French life-sciences corpus.

Low-resource languages and domains have long been constrained by scarce native corpora; the author team claims this route bypasses that bottleneck, and has released the TransCorpus toolkit, the pretrained TransBERT-bio-fr model, and reproducible code.

All performance conclusions come from the authors' own evaluations (self-reported), on French life-sciences downstream tasks, with no third-party comparison numbers; accepted to EMNLP 2025 Findings, with no independent reproduction yet — for most readers this is a method signal for low-resource language/domain modeling to track.

Tether AI Research released open-source translation models on September 2 — QVAC TranslatePsy-AfriSLM and EuroNano — that run fully offline on phones and laptops, keeping data on the device.

AfriSLM covers 19 African languages including Hausa, Amharic, Yoruba and Swahili, which Tether says reach roughly half of Africa's population; EuroNano uses English as a pivot for 9 European languages across 90 translation directions, with the smallest deployment at just 36MB. Models are downloadable on Hugging Face in 0.8B, 2B and 4B sizes.

All performance figures are Tether's own: the 0.8B version reportedly beats Qwen3.5-122B-A10B and other larger models on FLORES-200, BOUQuET and SMOL, attributed to filtering out about 96% of low-quality training data. These comparisons are not independently reproduced; the underlying research has been accepted at EMNLP 2026, a future verification point.

Vambo AI has released MORENA, a 1.5B-parameter model trained from scratch for 12 African languages, with self-reported benchmark results ahead of 26 tested models.

The developers report 1.408 bpb on African language modelling, the best of 26 models tested; the closest rival, Lugha-Llama-8B, carries over five times the parameters and scored 1.423 bpb. All figures are developer-reported, relayed via TechRadar, with no independent replication yet.

The mechanism difference is a custom vocabulary: African text encodes with 1.39 times fewer tokens than Gemma 3, though it still costs about 6% more tokens per byte than English, a gap the team says it cannot fully explain. Training took roughly 22,000 A100 GPU hours, with compute support from UNDP and others.

Microsoft opened the public preview of its security operations center (ISOC) on September 23, folding the log-and-alert analysis (SIEM) features of its analytics platform Sentinel directly into Defender.

The preview is available to customers with Microsoft Defender Suite, Microsoft 365 E5 or E7 licenses; organizations already running an active Sentinel workspace are excluded for now. Case management and workbooks work without setup, while user and entity behavior analytics and third-party data ingestion require extra configuration, and ingestion charges may apply.

The design gives security agents the same signals and controls as a human analyst, letting them investigate and act on incidents, with humans still approving high-stakes actions. Microsoft has not disclosed pricing, and Defender data is kept for 30 days at no charge during the preview. The claim that attackers are putting agents to work comes from Microsoft's own executives, not independent verification.