X Paper READ

Tibo announced on October 6 that Auto-review is free for users signed in through ChatGPT and does not consume their plan allowance. That extends his 28-day improvement pledge; universal availability and billing remain his announcement, rather than an account-by-account finding.

The feature reviews eligible requests to cross the coding agent’s sandbox boundary. A separate agent examines the proposed action and its relationship to the user’s task. A rejected request can lead to a safer alternative or human confirmation. This is permission review, with sandbox limits retained; it is not a review of every code change.

OpenAI described the mechanism in an April 30 research post, so the October 6 change concerns access and charging. Tibo’s clarified path is Settings → General → Permissions → Auto-review. Official documentation says it depends on an approval boundary; full-access or never-ask configurations do not provide the same review requests.

Fewer approval interruptions could ease decision fatigue. Incorrect approvals and refusals still matter, and automated review provides no blanket safety guarantee. We have not tested billing or measured its practical error rate.

Liz Ann Sonders compared September manufacturing and services surveys on October 6. ISM services PMI eased to 54.9 from 55.4; manufacturing slipped to 54.5 from 54.6. Both remained above 50. The services report was released October 5 and describes September.

Services new orders fell to 59.8 from 60.9 and business activity to 56.5 from 61.7. Employment rose from 47.8 to 50.1, returning to expansion after two contracting months. Prices increased to 74 from 72.6; supplier deliveries rose to 53.2 from 51.3, indicating slower deliveries.

These diffusion indexes measure the breadth of reported changes. They do not mean output grew 54.9% or inflation reached 74%. Expansion, better hiring and persistent supply pressure coexist here. The survey alone cannot determine interest-rate decisions or market returns.

The Nobel organization announced October 6 that Francis Halzen receives the 2026 physics prize for decisive contributions to IceCube and high-energy astrophysical neutrinos. Followed source ChineseWSJ reported the award.

IceCube uses roughly one cubic kilometre of Antarctic ice. Sensors register light produced when neutrinos rarely interact with matter. Their weak interactions demand a huge detector, but also allow them to carry information about energetic cosmic processes, complementing observations made with light.

Halzen proposed the concept in 1988 and the observatory was completed in 2011. The October 6 development is the award for that long effort. It does not mean the instrument has just opened or every detected neutrino has an identified source.

Brazil’s electoral court, TSE, confirms that Flávio Bolsonaro and Luiz Inácio Lula da Silva face a presidential runoff on October 25. This combines earlier captured reports as a results follow-up: leading round one is not election to the presidency.

At the court’s stated checkpoint—00:11 local time on October 5, with 99.99% of ballot boxes counted—Bolsonaro had 47.03% of valid votes and Lula 45.16%. Neither exceeded the majority needed to win outright. These are timestamped announcement figures, not a fresh poll or a final-result readback.

The separately labelled valid-vote total conflicts with the percentages and blank/null-vote figures, so we omit it. The morning check returned the same announcement checkpoint; the results portal exposed only its application shell, not readable final counts. These figures remain provisional. Unsupported fraud or foreign-influence claims are not included.

New York’s October 5 Executive Order 65 declares a statewide disaster emergency through November 4. It cites 108 cases reported during 2026 as of October 3, including 92 since July 15 in under-immunized rural communities across 18 counties. These are not one day’s new cases.

The order expands MMR vaccination and testing capacity under specified conditions. Eligible advanced emergency providers may vaccinate under non-patient-specific orders; pharmacists may vaccinate children aged two and older under such orders. Midwives face supervision and qualification conditions. Registered nurses receive specified sampling and testing authority.

Vaccinations must be registered within 72 hours. The order explicitly retains patient or legally authorized consent. A statewide emergency defines government powers and resources; it does not establish equal infection risk everywhere.

Liz Ann Sonders revisited the third-quarter CFO Survey on October 6. Duke University and the Richmond and Atlanta Feds surveyed 517 financial executives August 17–September 4 and released results September 23. Larger firms’ improving optimism offset declines among smaller and financially constrained firms.

Economy optimism eased from 60.6 to 60.3; own-firm optimism fell from 70.7 to 69.7, on a 0–100 scale. About 20% of small firms and 10% of large firms reported financing constraints on costs or opportunities. A small average change therefore masks differing conditions.

Expected 2026 selling-price growth rose from 4.7% to 5.3%. That forecast is sales-weighted and winsorized; optimism uses unweighted averages. Forecasts are not realized inflation, and this is an older survey. It offers context for September ISM readings, without establishing a causal link between different samples and periods.

Simon Willison’s October 4 experiment, shared on X October 6 Beijing time, asks a local four-bit Qwen3.8-27B model to express integer sums only in English words. Correct formatting is distinct from correct arithmetic.

A non-reasoning sweep scored 1,195/5,070, or 23.57%, across digit-length pairs from one to thirteen. A separate medium-reasoning run scored 167/169, or 98.82%. Different samples prevent treating those percentages as a controlled comparison.

On the same frozen 169 questions, non-reasoning scored 45 and reasoning 167. However, reasoning strength, output allowance and execution order changed: the non-reasoning run allowed 128 output tokens; the reasoning run allowed longer completion. This compares configurations rather than isolating a causal reasoning effect.

Median latency rose from 1.50 to 27.62 seconds and median completion tokens from 14 to 313. The setup used DGX Spark and Qwen3.8-27B-Q4_K_M. Those costs and conditions bound the result; we have not rerun it or generalized it to other tasks.

Gorden Sun introduced Strata October 4. This background item completes an earlier unresolved source; it is not a new October 6 release. The project runs Qwen3.8-Flash-Next through coordinated GPU, system-memory and storage use.

125B refers to the model’s parameter count; 12GB is the minimum graphics-memory requirement. Frequently used experts remain on the GPU while compressed experts reside in RAM and the CPU participates in inference. Alongside at least 12GB VRAM, the project specifies at least 32GB RAM, roughly 80GB disk space and an SSD recommendation.

The 32GB Coder variant removes about half the experts; the project warns of weaker non-code and CJK performance. A 48GB configuration supports smaller compressed files, while 64GB supports a wider selection. The headline does not promise the same complete model on every 12GB-GPU computer.

Initial loading may take one to three minutes and temporarily reduce responsiveness. Published speeds depend on hardware, quantization and engine version. We have not installed or benchmarked it. Match those conditions to the intended workload before judging practical value.

The morning documentation check is pinned to October 6 commit 82f46a8c. It describes that version, not necessarily the configuration available when the October 4 post appeared.

After recounting an AI-coding interview October 4, Bao Yu examined whether skipping individual PR reviews is reproducible. This article reports his own conditions. We read both posts, but have not watched the interview or verified its productivity figures.

First, he treats capable models and adequate token budgets as prerequisites; his personal multi-account practice is not a universal purchasing prescription. Second, AI-assisted checking still needs product observation and human judgment. Repeated deterministic checks belong in scripts.

Third, engineering direction should divide work into small, tested milestones. His marketplace example separates listings, browsing, user roles, payments and security before scaling. Fourth, derive tools and skills from repeated corrections and interventions in one’s own conversations, rather than copying another person’s setup.

The argument moves responsibility toward decomposition, verification and acceptance; it does not remove responsibility. These are practical recommendations, not a controlled demonstration that skipping PR reviews is safer. Risk and testability still shape the appropriate review process.

Dylan Patel shared SemiAnalysis’s subscription-limit analysis. Its public methods separate plan, model and token type, then value the available allowance at API list prices. That is an API-equivalent cost measure, not cash, refunds or completed-work quality.

The team isolates fresh input, cache writes, cache reads and output. Random tags on long prompts avoid reusing one cache entry; repeated fixed prompts test cache reads. Long completions test output. Meter-based estimates discard partial steps and account for other token costs.

Five-hour, weekly and model-specific caps are distinct. The comparison uses the team’s September agent-workload mix. Different cache or output patterns can change rankings. Falling API prices can reduce equivalent value even when subscription token limits remain unchanged.

The roughly fivefold claim is conditional on model and workload. The article reports about 20% variation between some same-plan accounts, attributed to a limited provider test—not a universal difference. The final third-party comparison is paywalled and unread; we report the public method, without an independent reproduction or a complete subscription ranking.

At 07:26 Beijing time on October 7, Rohan Varma posted a one-week improvement list for dots, reposted by ChatGPT. OpenAI previously introduced dots as an always-on agent. This update addresses interruptions, presentation and connections rather than announcing a model.

Reported shipped fixes cover faster cloud browsing, duplicate phone notifications while using ChatGPT, web code blocks and attachments, reply pauses and sidebar loading. Safari and Firefox character rendering and stalled Outlook setup are also listed. Clipped approval buttons, Gmail previews obscuring approvals, and enterprise access on web and voice startup received fixes.

The roadmap is separate: better voice reliability; Codex folder management, full thread context and renaming; clearer ongoing-task organization; multiple computers; and improved Windows local control. These are not all available now. The list points to smoother handoffs in agent work, but the speed claims are the author’s report, not independently measured performance.

OpenAI announced its internal model’s mathematical release at 06:19 Beijing time on October 7. The repository lists 722 manuscripts in 372 related families. Roughly 4,000 denotes attempted problems, not 4,000 verified theorems.

A family can contain a principal result, supporting arguments, consequences or alternative proofs, so manuscript counts do not count independent breakthroughs. The release includes some Lean formalizations, ten reasoning summaries and compute disclosures. Lean checks formal proofs; the specific statement and assumptions still matter.

Verification stages vary, and unformalized results may contain problems. Revisions will preserve earlier versions. The model remains unreleased; OpenAI reports average compute equivalent to roughly three hours of ChatGPT Pro thinking per result. This report checks the catalogue and disclosure process, not individual proofs, and does not declare the Riemann hypothesis or other major conjectures solved.

At 07:34 Beijing time on October 7, Eric Topol asked whether statins prevent glaucoma, linking a British Journal of Ophthalmology study. It compares existing TriNetX records retrospectively, without randomly assigning medication.

After matching, the incidence analysis retained 210,701 people in each group, over 40 with ophthalmological records. At five years, new diagnoses numbered 4,771 (2.3%) among statin users and 5,491 (2.6%) among controls. The hazard ratio was 0.85 (95% confidence interval 0.82–0.89). The relative association is not a 15-percentage-point reduction; the crude proportions differ by about 0.3 points.

Matching cannot remove unrecorded confounding such as adherence and dose. Eye pressure, visual fields and optic-nerve images were unavailable; treatment escalation in a separate analysis does not directly measure progression. The authors call for prospective randomized trials. The morning discussion provides no new indication to start a cholesterol drug for glaucoma prevention.

a16z highlighted Valon on October 7 Beijing time. The company announced $150 million in funding at a $2.3 billion valuation on October 5, with Ribbit joining and a16z returning. This is a financing follow-up.

ValonOS handles mortgage data, payments and compliance workflows. The company reports over $200 million in contracted annual recurring revenue within six months, and contracts covering one in six outstanding U.S. mortgages. Contracts are not recognized revenue or completed migrations. ServiceMac and Carrington are separately named as live users. These are company disclosures, not independently audited figures.