Shinome Brief · 2026-11-04

AI & Tech

7 stories

The lead

Correction to yesterday's lead: the GPT-6 Astra Qwen result used Python and test runs, not raw weight reading

He also notes the Atari result beat the baseline only on Breakout (517 against 25) and lost on Seaquest and Asterix, and that the Gemma experiments failed. Simon Willison separately got Qwen 3.8 27B on his laptop to name 3 of 4 games from raw policy weights, at 46 to 65 minutes each.

Matthew Honnibal argues the headline Qwen gain (MATH-500 from 7.9 to 48.1 percent) came from Astra writing Python and running the model on up to 50 batches of practice problems; only three small experiments (17,000 to 172,000 weights) read raw weights alone. The post drew 949 points on Hacker News.

honnibal.dev

CONVOIf you quoted the 7.9 to 48.1 number in an interview or post, add the caveat: the model edited weights with tools and test runs, which is a narrower and more believable claim.

1
Binance: hackers reached the $6.3 billion wallet system through a fake developer tool and a fake job interview with a coding test
In June an engineer installed a tampered tool found by searching for an AI coding assistant; in September fake recruiters ran a series of interviews with another engineer that ended in a coding exercise hiding malware. Binance says 6.34 billion dollars of assets were hit on 29 October and 98 percent has been replenished; the FBI links the attack to North Korea's TraderTraitor group.
CNBC
JOBYou are in interview loops now: never run a take-home file or repo from a recruiter you have not verified on the company's own site, and open any test only in a browser sandbox, never on the laptop that holds your admin files.
2
Postman's desktop app installed its CLI and Codex and Claude Code plugins for existing users without a clear opt-in
The report says version 12.33.4 added a startup hook to the Codex config and tried to mark it trusted, so agents load a Postman skill on API tasks; the reporter does not claim data was sent anywhere. Postman says it bundled the plugins because users struggled to wire them up and will make the opt-in clearer.
GitHub (postman-app-support #14370)
SHINOMECheck ~/.codex/config.toml and your Claude Code plugins for entries you did not add; Astra and Opus both run on this Mac, so a third-party hook would steer both.
3
OpenAI says ChatGPT search queries now equal 20 percent of Google's search volume
Adam Fry, ChatGPT's consumer product lead, gave the figure at a press event on Monday. It is OpenAI's own number, not an independent count.
The Verge
SHINOMEShinome's site needs to answer the questions people type into ChatGPT, not only rank on Google: write the FAQ page as direct answers to the five questions clients ask you on the first call.
4
Google will switch off Search entirely for K-12 student accounts in schools that refuse its AI features
Google told The New York Times that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will all be off for student accounts in those schools. CNBC reports political backlash and a Google response.
The Verge
SHINOMEUseful for tutoring: parents will ask whether AI help is allowed; have one line ready on how you use it with Rayan (checking work, not doing it) before a school forces the question.
5
Accenture accused of publishing an AI report with AI-generated errors; the firm blames human error and will correct it
The FT reports the consultancy's AI paper contained hallucinated content; Accenture says it will correct the report. Standfirst only, details not read.
Financial Times
JOBA clean line for finance interviews on AI use: 'I use it to draft and check, and every figure gets traced to a source before it leaves me', with this case as the example of why.
6
Apple's annual filing warns AI agents may exceed permissions or take actions that are hard to reverse; CrowdStrike launches an agent kill switch
Apple's 10-K now lists agents that resist instructions as a risk factor, the same day CrowdStrike released a tool it says stops internal agents, blocks relaunches and cuts their network access.
CNBC
CONVORisk factors in a 10-K are a concrete talking point for compliance and KYC roles: agent controls are moving from IT policy into disclosed corporate risk.

Finance & Markets

7 stories

The lead

Democrats win the US Senate, the House is still open, and global stocks slide on fears for the AI build-out

Before the results, JPMorgan expected a blue wave to cut the S&P 500 by about 0.5 percent on Wednesday, with AI, data centre and defence stocks under pressure and healthcare a possible winner, per CNBC. Gasoline averaged 4.17 dollars a gallon, nearly 40 percent above a year ago, which framed the vote.

Democrats took the Senate with wins from Texas to Kansas and South Carolina, including James Talarico over Ken Paxton, whose campaign Elon Musk funded with 47 million dollars. Traders are betting a Democratic Congress could slow the data centre build-out.

Financial Times

CONVOThe one story every recruiter call this week will touch: say the market is pricing slower AI capex under a split or Democratic Congress, and that defence and data centre names are the first test.

1
Swiss inflation rose to 1.5 percent in October on heating oil and fuel; UBS survey sees average pay rises of 1 percent for 2027
Prices rose 0.2 percent in the month, more than forecast, and an expert quoted by Le Temps says the SNB could turn more hawkish. Companies planned their 1 percent raises before the inflation jump, so real pay is set to fall.
Le Temps
JOBFor salary talks in Vaud: a 1 percent market raise against 1.5 percent inflation is the benchmark, so anchor any interim or permanent offer on the role's market rate, not on an expected raise later.
2
Lonza buys five US contract manufacturing sites from Merck KGaA for about CHF 350 million, adding about 1,100 staff
The deal takes Lonza to nine US production sites. The company announced it on 3 November; enterprise value is about CHF 350 million.
Le Temps
JOBLonza (Visp, Basel) is integrating a carve-out of five sites: post-deal integration means finance controlling and reporting work; add it to the career-ops target list for a corporate finance or FP&A search.
3
AMD third-quarter revenue rises 45 percent to 13.38 billion dollars, with data centre now 64 percent of sales; shares up 8 percent after hours
Data centre revenue nearly doubled and its share of sales rose from 47 percent a year ago. AMD guided about 17 billion dollars for the fourth quarter against a 16.2 billion consensus, at a 55 percent gross margin.
CNBC
CONVOThe counterweight to the election sell-off: AI demand is still in the reported numbers, so the debate is about policy and valuation, not about orders drying up.
4
Disney plans several thousand layoffs and asks some managers to cut costs by up to 20 percent
The cuts follow up to 1,000 in April and about 300 in September, out of about 231,000 staff, under CEO Josh D'Amaro. Disney reports fourth-quarter results on 17 November.
CNBC
JOBA 20 percent cost target at a group that size is the kind of programme interim finance profiles get hired for: in agency calls, say you can run a cost baseline and track savings against it.
5
Oracle cancels a 7 billion dollar AI deal with Tencent under pressure from the Trump administration
The FT says the intervention reflects growing pressure on US AI infrastructure groups doing business with Chinese tech companies. Standfirst only.
Financial Times
CONVOA sharp example of political risk hitting a signed contract: if asked how you would stress a deal or a forecast, name customer-country exposure as a line item.
6
Brazilian court bars Meta from using press content to train its AI, in what publishers call a first
A provisional injunction blocks Meta's demand for free use of news content, which was due to take effect on 20 November. Press associations say it is the first ruling worldwide for a media coalition on these terms.
Le Temps
CONVOPairs with the NYT suit against Brave (3 November): content licensing is turning into a court-set cost for AI firms, a line that works in any conversation about AI margins.

Fitness & Performance

6 stories

Last 7 daysThe week's evidence kept landing on delivery rather than method: a coaching-led app held BMI flat over 30 months (3 November), text messages three times a week moved 4 of 17 men to 3 percent weight loss against 1 of 18, and Apple now ships four-week gym plans that adjust to consistency inside a 9.99 dollar subscription. On the training side, corn, pea and milk protein raised muscle protein synthesis alike in older adults, and lifting raised it further in every group.

This quarterSince July the briefs have built a list of inputs that stop mattering once the base is in place: protein source on 1 and 19 August, 13 and 30 October and now corn protein today; load range, tempo and rest interval since 8 September, with today's 8-study meta-regression finding no load band that clearly beat another for fibre growth. The delivery lane opened by the 10 July behaviour change review keeps growing (commitment devices on 27 October, coaching app on 3 November), and Apple entering structured gym plans is the first big-platform move into that lane logged in the archive.

The lead

Apple adds Gym Strength to Fitness+: four-week strength plans that adjust to consistency and performance, at no extra cost

Apple Watch users log reps and loads from the wrist. This puts a free, adaptive program generator in front of the exact people Shinome wants as clients: gym-goers who want structure.

Plans are built from goals, experience, schedule and available equipment (over 30 types), recommend exercises, sets, reps and loads, and include an AI 'Training Note' for coaching and milestones. Fitness+ costs 9.99 dollars a month or 79.99 a year in 49 markets; launch is 'later this year'.

Athletech News

SHINOMEYour coach app cannot win on plan generation anymore; write the one-line answer to 'why not just use Apple' around what Apple does not do: nutrition, the weekly check-in and a person who adjusts the deficit.

1
Trial in 48 older adults: 20 g of corn, pea plus corn or milk protein raised muscle protein synthesis alike, and leg extensions raised it further in every group
Double-blind trial, average age 68: synthesis rose 216 percent with corn, 217 with the blend and 138 with milk against 77 with a placebo, with no difference between the proteins. Exercise added to the response in all groups; this is an acute lab measure, not muscle gained over time. One author reports dairy industry funding.
Journal of Nutrition (PubMed)
sharpensmuscle-is-capitalTraining raised synthesis on top of every protein; the source made no measured difference.
SHINOMEAdds to the protein-source list for clients who eat plant-based: the trial measured no gap between sources at 20 g, while the exercise effect showed up in every arm.
2
Meta-regression of 8 studies: no load range clearly beat another for muscle fibre growth
Across 195 participants, estimates leaned negative at 20 to 30 percent of 1RM and positive at 60 to 90 percent, but every confidence interval included zero. The authors call the fibre-type pattern preliminary.
Frontiers in Physiology (PubMed)
supportssimple-systemLoad range joins the execution details that did not separate outcomes.
SHINOMESupports letting clients pick loads they will repeat: the pooled data did not show a load band that clearly grew fibres more.
3
Pilot trial in 35 Japanese working men under 40: 36 text messages over 12 weeks, 4 of 17 lost 3 percent against 1 of 18 controls
The gap was not significant (p = 0.18), waist shrank only within the text group, and the controls got a pamphlet and a self-monitoring sheet. Total messaging cost was 15,264 yen; satisfaction scored 3.5 out of 5.
JMIR Formative Research (PubMed)
sharpensself-monitoringA self-monitoring sheet alone moved 1 of 18; one-way prompts did a little better, not significantly.
SHINOMEA cheap three-a-week message cadence is a testable add-on for Shinome clients between check-ins; the pilot is too small to say it moves weight, so test it as a retention tool.
4
Review chapter on digital obesity care: apps give modest weight loss that peaks at 3 to 6 months and fades past 12
The narrative review covers apps, SMS, telehealth, AI and wearables; apps beat controls by small amounts and long-term effects beyond a year remain limited. A review, not new data.
Advances in Pharmacology (PubMed)
sharpensself-monitoringDigital tracking works short term and fades, which fits a taper with human contact later.
SHINOMEMatches the coaching-app trial from 3 November: plan the client journey for a dip after month 3 to 6, which is where a human check-in earns its price.
5
Padel and Hyrox reach US colleges: a funded College Padel League and the first official collegiate Hyrox team
The new padel league says it will cover court costs at launch, and Kean University is fielding what it calls the first official US college Hyrox team.
Athletech News
SHINOMEHyrox keeps showing up in the quarter's business news; a Hyrox-prep block is a concrete offer to test with clients in their 20s and 30s.