Shinome Brief

Saturday 7 November 2026

Private edition · 11 min read

Web search returned stale results again; built from feeds plus primary pages (Ars, CNBC, Google, Crusoe, Hightouch, Planned Obsolescence, PubMed eutils). Le Temps and FT bodies blocked, those items use feed text. Fitness business is thin: Athletech's Peloton, Planet Fitness and NYC Marathon pieces are paywalled with no numbers.

AI & Tech

6 stories

The lead

Chainalysis says names, usernames, password hashes and live tokens for all its users were exposed in an attack that used AI agents

Its customer list includes the FBI, ICE, the US Air Force, Microsoft, Coinbase and Barclays. The company found the exposure on Monday but disclosed it on Wednesday, and has said nothing about how the AI agents worked.

Chainalysis first called the October breach a leak of non-sensitive data, then said on Wednesday that an on-path attack had exposed credentials for every user. It took its login page and its KYT transaction monitoring offline; customers now need allow-listed IPs and a mandatory credential reset to get back in.

Ars Technica

JOBChainalysis KYT and Address Screening are standard compliance tools; for KYC and compliance interviews, have the line ready that the screening vendor itself was down this week and every bank user had to rotate credentials, so vendor outage plans are part of the job.

  1. 1

    Ajeya Cotra reads METR's report as saying AI now drives about 80 percent of a doubling in Anthropic's research pace

    From October 2025 to October 2026 Anthropic's capability progress ran about 2x its prior pace on an internal index, and METR credits roughly 80 percent of that to AI, a central 1.8x speedup. Cotra, who did not work on the report, notes much of the evidence is redacted.

    Planned Obsolescence

    CONVOWhen AI comes up with a senior person, the sharp version is: the debate moved from whether AI helps coders to whether it is speeding up its own labs, and the main evidence is a partly redacted METR report.

  2. 2

    Hightouch: a cheap model that hands its unsure cases to Claude Opus matched Opus accuracy at about a quarter of the cost

    On 1,000 identity-matching checks, Opus alone scored 99.0 percent for 3.86 dollars, and the small Jev model escalating its 30 to 90 percent confidence cases to Opus scored 98.8 percent for 0.88 dollars. On the harder set the hybrid matched Opus (95.3 against 95.1) at 1.8x lower cost, because the small model's confidence was well calibrated.

    Hightouch

    JOBFor data and finance ops roles, this is a clean answer to how you would cut AI costs: route everything to a cheap model, send only the uncertain middle to the expensive one, and pick the cheap model on calibration, not raw accuracy.

  3. 3

    Swiss firms move to open-source models, often Chinese, as AI bills grow, Le Temps reports

    From Apple to Revolut, large tech users are rationalising AI spend, and Swiss companies are now choosing open-source models, many developed in China, to control costs. The full article is paywalled; the figures it cites could not be read.

    Le Temps

    CONVOA Swiss angle for interviews on AI budgets: cost, not capability, is now driving model choice in Swiss companies, and that raises the data-residency and compliance questions a finance or KYC hire gets asked.

  4. 4

    Google ships Gemini Omni 1.2 Flash: video clips of 3 to 15 seconds at 0.07 dollars a second in 720p, 30 percent cheaper

    The model edits clips turn by turn without regenerating them, follows character and product reference images more closely, and drafts at 360p or 720p before upscaling to 4K. Google Vids users get 15-second clips at no extra cost.

    Google

    SHINOMEFor short exercise and brand clips, a 10-second 720p clip costs about 0.70 dollars, so test one cue video in Google Vids before paying anyone to film it.

  5. 5

    Google buys Crusoe's planned data centre site in Wyoming after Crusoe finished the permits and power plan

    Crusoe did site selection, approvals and a hybrid power plan with Tallgrass Energy (fuel cells, grid and gas turbines), then sold the site to Google. Crusoe says it has over 6 GW contracted and 40 GW in its pipeline.

    Crusoe

    CONVOPair it with Microsoft trimming its neocloud partners in Finance below: the hyperscalers are buying de-risked sites and dropping middlemen, which is the shift to name when data centre debt comes up.

Finance & Markets

6 stories

The lead

US adds 165,000 jobs in October and unemployment falls to 4.1 percent, but the average jobless spell reaches 27.1 weeks

CNBC ties part of it to cuts at Disney and a 5 percent AI-linked job cut at American Express. The pattern is low firing, slow hiring: people who lose a job take longer to find the next one.

Payrolls beat forecasts, and the FT reports markets now expect the Fed to raise rates. Under the headline, the average unemployment spell is the longest outside the pandemic since 2016, and 27.6 percent of the jobless, about 2 million people, have been out of work for over six months, the highest share since late 2021.

CNBC

JOBUseful framing for his own search and for recruiter calls: hiring is slow everywhere, not just for him, so the edge is speed and volume through interim agencies rather than waiting on one permanent role.

  1. 1

    Anthropic presses ahead with an IPO before Thanksgiving, but investors now price it at 1 to 1.5 trillion dollars; OpenAI closes 11 billion at 1.25 trillion

    Anthropic, valued at 965 billion in May and once aiming for up to 2 trillion, plans to list on Nasdaq as ANT with the roadshow starting next week; one investor sees 1 to 1.25 trillion as more likely. OpenAI raised 11 billion at 1.25 trillion, below the 30 billion at 1.4 trillion it first sought (Bloomberg via The Verge).

    CNBC

    CONVOThe one-line take for a senior room: both labs are still raising, but at lower prices than they wanted a month ago, so the AI rout has reached private valuations, not just listed stocks.

  2. 2

    CoreWeave's 2032 bond yields over 16 percent and its credit default swaps topped 1,289 basis points as AI debt comes under strain

    Goldman counts 67 billion dollars of AI bond deals since 29 October and says data centre moratoriums in Democratic-led states could touch 45 percent of planned US capacity. Nvidia's CDS rose 76 percent in a month, from a low base, and hyperscalers ran negative free cash flow in the third quarter.

    CNBC

    CONVOA corporate finance talking point with numbers: the AI build-out is now a credit story, and a 9.625 percent coupon trading above 16 percent tells you the market prices CoreWeave close to distressed.

  3. 3

    Microsoft cuts back on neocloud partners: it did not renew deals with Oracle and Nscale, and is still deciding on others

    Microsoft spent 45.2 billion dollars on capex and finance leases last quarter and is focusing on fewer key suppliers. CoreWeave got 62 percent of its 2024 revenue from Microsoft; CoreWeave, Nebius and IREN shares are down about half over a month.

    CNBC

    CONVOA customer concentration case to cite in a finance interview: one client at 62 percent of revenue, and that client is now consolidating suppliers.

  4. 4

    Lagarde's ECB succession comes down to the Netherlands' Klaas Knot and Spain's Pablo Hernandez de Cos

    They are the two official candidates for the ECB presidency from February, and eurozone finance ministers review them from Monday, with inflation pressure in the background.

    Le Temps

    CONVOKnow the two names before a Swiss bank or treasury interview: the next ECB president sets euro rates, which feed straight into the franc and the SNB's choices in 2027.

  5. 5

    Washington tells Beijing to meet the Busan trade deal terms on rare earths immediately

    The US says China is not raising supply of rare earths and other critical products as it committed. The article body is paywalled; this is the FT's own summary.

    Financial Times

    CONVOSwiss industrials such as ABB, already hit by China blocking its robotics sale this week, are the local angle if trade comes up.

Fitness & Performance

5 stories
Last 7 daysThe week's science kept rating evidence more than finding new effects: the 5 November umbrella review called 30 of 33 reviews critically low confidence, today's frailty meta-analysis found strength gains but only low to very low certainty, and a radiofrequency pilot added diet to a device but left out the diet-only arm that would test the device. Most of what was published this week is too small or too loose to change practice.
This quarterResistance training in older adults is the longest thread in the archive (8 August and 5 September reviews, the age-blunted gains of 13 October, FAST-2 on 21 October, the 28 October quarter review of combined training), and the result has stayed the same: strength improves, frailty and function scores stay uncertain. trends.py logs energy-balance challenged once in 35 tags, and the 21 August review rejected the carbohydrate-insulin model; today's protein partitioning paper is a third model, built on theory with no new trial data.

The lead

Meta-analysis of 9 trials in 566 frail and pre-frail older adults: lifting raised knee strength, but frailty scores stayed uncertain

Published 22 October and indexed this week, it only counted trials of structured resistance training with no other exercise added. The authors call for larger trials with standard programmes and longer follow-up.

Resistance training alone, against controls, improved knee extensor strength (SMD 1.30, low certainty) and possibly handgrip (4.15 kg) and gait speed (0.13 m/s, both very low certainty). Effects on Timed Up and Go, SPPB and frailty score were uncertain, and heterogeneity was substantial.

Frontiers in Public Health

sharpensmuscle-is-capitalStrength gains held in the frailest group, but the link from strength to function and frailty scores was not shown with confidence.

SHINOMEFor older clients or a parent's question, the accurate claim is narrow: lifting made frail adults stronger in the legs, while proof that it lowers frailty itself is still missing.

  1. 1

    A theory paper proposes a 'protein partitioning model' in which protein plus lifting shapes body composition at the same calories

    The authors combine a muscle-centred model with an economics model of habit to argue that protein with resistance exercise settles at lower body fat, while carbohydrate-led eating swings more. It is a model only: no participants, no trial, no measured outcome.

    Theory in Biosciences

    challengesenergy-balanceArgues energy partitioning, not just the calorie total, drives composition, but offers theory and no data.

    SHINOMEIf a client brings this framing, the honest line is that it is a hypothesis paper; no study has yet shown composition changes at matched calories the way it predicts.

  2. 2

    Pilot trial in 20 adults: radiofrequency fat device plus a low-calorie diet beat the device alone for weight and BMI

    Over 6 weeks and 12 sessions, the diet group lost 15.0 percent visceral and 18.6 percent subcutaneous fat against about 9 percent with the device alone, but the fat differences were not statistically significant; only weight and BMI differed. There was no diet-only or sham group.

    Lasers in Medical Science

    supportsenergy-balanceAdding a calorie deficit produced the only significant group difference; the device's own effect was not tested.

    SHINOMEWhen a client asks about fat-melting machines, this is a clean example: the only clear between-group difference came from adding the diet, and the device was never tested against no device.

  3. 3

    Observational study in 46 children aged 7 to 10: more ultra-processed food tracked stronger brain reward responses to food pictures and more body fat a year later

    All children started at a healthy weight; higher dorsal striatum activity on MRI went with larger rises in BMI and body fat. Adjusting for that response weakened the food-to-BMI link by 31 percent and the food-to-body-fat link by 44 percent; the design cannot show cause.

    ScienceDaily (JAMA Network Open)

    SHINOMEBackground for tutoring parents or Shinome content on kids' food: report it as a link in 46 children, not as proof that these foods change the brain.

  4. 4

    Hybrid race affiliations cost gyms 130 to 159 dollars a month, against 4,500 a year for CrossFit, and single events net 5,000 to 20,000

    Over 2 million people are expected to race Hyrox this year. A Hyrox affiliate in Newfoundland went from 182 to 330 members, one Hyrox camp grossed 30,000 dollars, and a Texas DEKA event drew 372 people for 12,000 dollars of profit in a day; Deadly Dozen lets gyms keep all registration fees.

    Athletech News

    SHINOMEA pricing benchmark for Shinome: a one-day race simulation earns gyms 7,000 to 9,000 dollars, so a paid group test-day for clients is a revenue idea worth costing, not just a community event.