독서

Tesla has released an "Emergency Drive Away" feature that lets the car shift into drive while the charging cable is still connected. Until now, a Tesla would not shift out of park when a cable was detected.

The feature arrives via over-the-air update: the driver shifts into drive and taps an "Activate" button, and the charging handle breaks off. Tesla said on X that using it will damage the vehicle and the Supercharger, and that deliberate misuse will invoke additional penalties — but the company did not say what those penalties are.

The rollout follows a shooting in Twin Falls, Idaho, two months earlier that left three people dead, one of them inside a Tesla at a charging station. Max de Zegher, Tesla's senior director of charging in North America, called it a feature "we hope you never need," saying the team went back and forth on it for years.

Singapore's Competition and Consumer Commission has taken action against 45 companies that bought fake five-star reviews, its biggest crackdown on the practice so far.

The companies — including a funeral parlour, a law firm and an aesthetics clinic — bought reviews from a provider whose services included AI-generated testimonials resembling real customer experiences. About 100 companies were involved; in the first phase the 45 had to take down the reviews and publish a public apology for six months.

The enforcement details come from SCMP reporting; the commission's own announcement was not in the material, so the legal basis and later phases remain to be confirmed.

Per diginomica's October 1 report, Accenture CEO Julie Sweet argues that the more token costs fall, the more companies will deploy AI at scale — and the process re-invention that requires is Accenture's business.

On the numbers: fiscal Q4 revenue was $18.7 billion, up six percent year on year, with profit of $1.99 billion; full-year revenue reached $74.2 billion. Nearly 100 additional clients started their first advanced AI work in the quarter, bringing the fiscal 2026 total to more than 400.

Note that the falling-tokens-boost-demand logic is the company's own framing, and Sweet said a clearer picture of overall client budgets should come in January and February.

China's quantitative investing sector has added 18 new funds managing more than 10 billion yuan (US$1.5 billion) this year, bringing the total to 159, according to financial news agency Cailian Press as reported by the South China Morning Post on October 2.

New entrants include AXQ Capital, Hopeseek Fund and Huanian Fund, per financial data provider Wind. The report attributes the growth to technology-driven momentum in Chinese equities.

The sector remains under tight regulatory scrutiny. The "DeepSeek effect" framing is the outlet's interpretation, not part of the underlying tally.

From October 1, 2026, Hubei province expanded its trade-in subsidy program: newly covered categories include embodied robots, drones and blood-pressure monitors, subsidized at 15% of the final sale price.

The list covers smart cleaning devices, service robots (including cooking robots, exoskeletons, companion robots and robot dogs), smart earbuds, locks, kitchen appliances, toilets, digital cameras (including action cameras and drones), products for the elderly, and smart furniture. Each consumer may claim one item per category, capped at 1,500 yuan per item.

The adjustment implements the eight-ministry implementation opinion on "AI plus consumption." According to IT Home's October 2 report, the original policy document was not linked, and the list of participating merchants is subject to official Hubei announcements.

After a one-minute video call, 48 percent of participants believed Tavus's AI avatar Griffin was a real person — the result of Tavus's own study.

Previous systems topped out at a two percent mistake rate in comparable tests, nowhere near this level of realism. Griffin is a real-time video conversation model from the San Francisco startup, released around September 30, processing speech, facial expressions, tone and gestures while generating video, per The Decoder's October 1 report.

Tavus also cites a test it describes as independently run by Nvidia: Griffin scored 3.83 on how human it feels, close to 3.92 for actual humans, with the previous best AI at 2.80. Only a preview called Griffin-Lite is available to select testers; the fuller version waits until safety concerns are addressed.

Masked self-distillation internalizes the chain of thought into model parameters: on Qwen3-4B and Qwen3-8B, the authors' own tests report task performance held or improved while the model generates fewer intermediate steps, cutting inference latency and cost.

Previously, a chain of thought is the long sequence of explicit reasoning steps a large reasoning model produces before its final answer, and these traces dominate serving latency and compute. The method instantiates copies of the same model as teacher and student, training the student to internalize part or all of the intermediate trace and emit shorter outputs.

The authors ran controlled experiments in two domains, math and graph coloring, reporting no catastrophic forgetting out of domain; an ablation found plain supervised fine-tuning also shortens traces but generalizes worse. All results are the authors' own, and specific numbers require reading the full text; the preprint was first submitted June 18 and revised to v3 on September 30, with no independent replication yet.

Apple's machine learning research team reports that strengthening a model's ability to discriminate languages during pretraining narrows the gap between multilingual speech models and monolingual ones.

In a controlled English-French HuBERT setup, adding an auxiliary language classifier or per-language clustering targets cut phone discrimination error from 11.6% to 10.4%, below the monolingual model's 10.8%.

Lexical performance rose from 52.1% to 56.7%, still behind the monolingual 58.5%. These are the authors' own controlled experiments, and the findings are limited to the English-French setting.

On October 1, Ecuadorean authorities detained Swedish free software developer and security expert Ola Bini in Quito and ordered his immediate deportation, barring him from returning for 10 years.

The grounds are a secret report alleging acts against Ecuador's security, whose contents the defense could not access. His lawyers filed a habeas corpus petition. According to information released by his lawyer, officials showed no document supporting the visa revocation, and human rights organizations were denied entry to the deportation hearing.

The old case behind Bini's criminal conviction has expired under the statute of limitations, and the court had already formally lifted all precautionary measures. EFF has followed the case for years and supports Bini; the account above comes mainly from the organization relaying his lawyers' posts, though Ecuador's immigration office also published a statement on its official account.

Some prediction systems give a reason for each decision: this loan was rejected because income was the factor with the largest share. SHAP is an algorithm that sets the rules for this kind of explanation; it breaks down and computes each factor's contribution to the outcome.

The explanation bar charts behind today's risk-control and medical models are mostly powered by this algorithm. The 2017 paper proved that if you want reasons to satisfy a few common-sense properties — each factor states its own amount clearly, unclaimed portions are accounted for, and the reasons don't change depending on how factors are grouped — only one algorithm can do it. Later explanation tools are basically patches on top of it.

If the model is complex and factors interact with each other, exact computation is impossible, and approximations rely on extra assumptions — don't use it to judge contributions when interactions are strong. Small-scale user experiments are also not enough to prove that people actually trust its explanations more.

《A Unified Approach to Interpreting Model Predictions》(2017) | Next review 2027-09-20