Shinome Brief · 2026-09-07

AI & Tech

6 stories

The lead

OpenAI's chief scientist says no lab has solved alignment well enough to keep scaling at full speed

This is a step past the 6 September commitment to publish a misalignment reporting framework, which this brief led on yesterday. Yesterday OpenAI promised a process; today its chief scientist says the industry's main safety technique is losing reliability and that the company expects to slow down.

In an essay published 6 September, Jakub Pachocki writes that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer", says he expects and hopes voluntary slowdowns become commonplace, and calls for the Preparedness Framework and Anthropic's Responsible Scaling Policy to become mandated safety bars enforced by third party auditors, government agencies or international bodies. He also says OpenAI's own evaluations show its ability to rely on chain of thought monitoring is progressively diminishing.

OpenAI

CONVOThe line worth carrying into a recruiter or interview conversation: the lab shipping the fastest models is the one now asking for enforced limits, which reframes AI regulation from a Brussels story into an industry one. Pairs with the Astra release from 3 September if anyone brings up the pace of models.

1
OpenAI publishes internal data showing agent spend per researcher went from near zero to about 600 dollars a day in seven months
A companion post charts coding agent use inside OpenAI's own research team, with daily spend per researcher climbing from close to zero in February 2026 to roughly 600 dollars a day by late August 2026. Willison notes the acceleration in late July lines up with internal access to the model later shipped as GPT-6 Astra.
OpenAI, via Simon Willison
JOBUse this as the number when a recruiter asks what AI fluency means in practice: the people building the models now run agents at roughly 600 dollars of compute a day each, so "I use ChatGPT" is no longer an answer. It also sets the ceiling for what your own agent workflow is worth defending in an interview.
2
Professionals paid 37 dollars an hour to train the AI that replaces their own work
A report on the expert data market covering Outlier, Mercor and Surge finds South African professionals offered about 600 rand an hour, roughly 37 dollars, against a national minimum wage of 30.23 rand or about 2 dollars, while workers in India capturing video for robot training get about 250 rupees, roughly 2.60 dollars an hour. The tasks include teaching systems to mark essays and rewriting AI text to sound human.
Rest of World
JOBDirect read on the Mercor lane you planned to open in September: the pay gap is real but it is set by where you sit, and the highest rates go to people who bring a licensed or credentialed judgment, so lead with the corporate finance and compliance background rather than general writing.
3
The Seattle Times and Newsday sue OpenAI and Microsoft and ask for the models to be destroyed
Filed 4 September in the Southern District of New York, the complaint says both companies scraped the papers' sites including paywalled content to train ChatGPT, Copilot and Bing AI features, and reproduce the journalism verbatim. Beyond damages, the papers ask the court to order the impoundment or destruction of the training datasets and models that contain their work.
TechCrunch
CONVONote the contrast for any conversation about AI legal risk: on 3 September the Justice Department told a court that training on copyrighted text is fair use, and two days later publishers asked a court to destroy the models. The unresolved question is the reason enterprises still buy indemnities, which is a useful thing to know if you interview anywhere near procurement or risk.
4
Nitter and XCancel are back online two weeks after X's cease and desist
X Corp sent cease and desist letters on 24 August alleging API circumvention under the Texas Harmful Access by Computer Act and the Lanham Act, and both services went dark. On 6 September the maintainer said the project continues and service resumed after taking legal advice.
Hacker News
JOBPractical for your research stack: reading X threads without an account works again, which matters for the /last30days X lane and for checking what a hiring manager or a firm has posted before a call. Treat it as fragile and do not build anything that only works if it stays up.
5
AI labs shipped four model updates in one week and buyers report model fatigue
Anthropic, OpenAI, Meta and Google all released model updates in the same week that Nvidia said it was acquiring Hugging Face. The piece reports enterprise buyers struggling to justify re-evaluating and re-integrating on every release.
CNBC
CONVOA grounded contrarian line for a senior conversation: the constraint is no longer model quality, it is the cost of switching, so the firms winning are the ones that stopped chasing every release. Say it when someone asks whether you keep up with AI, because it shows judgement rather than enthusiasm.

Finance & Markets

6 stories

The lead

AfD takes 43.8 percent in Saxony-Anhalt and lands three seats short of governing a German state

The local branch of Germany's domestic intelligence service classifies the AfD as right wing extremist, so this is the first time a party under that label is within reach of leading a state since the war. Merz's personal approval is at record lows, and his economic reform programme is the thing voters rejected.

The AfD won 43.8 percent on Sunday, more than double its 2021 result, against 17.2 percent for Friedrich Merz's CDU. It holds 39 of 83 seats, three short of an absolute majority, and every mainstream party refuses to work with it, which leaves the state hard to govern.

CNBC

CONVOThis is the story senior people will open with on Monday. The finance angle worth having ready: Germany is the export market Swiss industrial suppliers depend on, and a government that cannot pass reform is why the Volkswagen restructuring from 4 September will not be the last one. Useful if you interview anywhere with German industrial exposure.

1
The 10 year Treasury faces a 4.8 percent test with 8.4 trillion dollars of debt to roll before year end
Matt Maley of Miller Tabak says a sustained move above 4.8 percent, the January 2025 high, would start creating meaningful problems for other asset classes. More than 8.4 trillion dollars of government securities roll over between now and year end, September may be a record month for high grade corporate issuance, and Treasury Secretary Scott Bessent's attempts to talk yields lower have not worked.
CNBC
CONVOKeep the level, not the narrative: 4.8 percent on the 10 year is the line to quote, and this brief had it at 4.78 percent on 4 September, so it is sitting on the threshold now. If a market conversation starts, that one number does more work than a view.
2
Beijing puts 360 billion yuan into three state banks and five insurers, with the tobacco monopoly helping to pay
The 360 billion yuan, about 53.6 billion dollars, comes through special treasury bonds led by the Ministry of Finance alongside China Tobacco, and is smaller than the market expected. It is the first time Beijing has extended recapitalisation to insurers, with China Life taking 35 billion yuan and China Taiping 7 billion, and PICC raising up to 15 billion through a private A share placement to the finance ministry.
Financial Times
CONVOWorth knowing as a slow risk story rather than a headline: when a state has to recapitalise its insurers, the stress has moved past the banks. Bring it up only if someone raises China, and stick to the fact that insurers are new to the list.
3
Pierre Mirabaud accepts two years suspended over 82 million francs of kickbacks to a Kuwaiti official
The former partner of Banque Mirabaud and past president of the Swiss Bankers Association, now 78, admitted having 82 million francs of hidden retrocessions paid between 2000 and 2012 to the director of Kuwait's social security, in exchange for mandates worth about 600 million dollars. The simplified procedure sets two years fully suspended for bribery of foreign public officials and aggravated money laundering, and the Federal Criminal Court validates it on 8 September.
Le Temps
JOBLive ammunition for interim compliance and KYC applications in Geneva and Vaud: this is the second retrocession case in a week after the 16 million franc Geneva prosecution on 3 September, and it is exactly why private banks are staffing remediation mandates. Name the pattern, not the person, when you write to agencies this week.
4
The Dutch central bank moved 86 tonnes of gold out of New York and Ottawa
Between March and August, De Nederlandsche Bank sold 59 tonnes held in New York and rebought in London, and physically shipped more than 27 tonnes, citing crisis preparedness and easier tradability at the Bank of England. London now holds 32.1 percent of its reserves, Zeist 30.8 percent and New York plus Ottawa 18.5 percent. France moved 129 tonnes out of the New York Fed between July 2025 and January 2026, booking an 11 billion euro gain.
CNBC
CONVOThird data point in the same direction after Norway's fund proposed cutting its US Treasury weight by a third on 4 September. The framing to use is boring and correct: reserve managers are diversifying custody and jurisdiction, not calling the end of the dollar. That distinction is what makes it sound senior.
5
The SEC sues ISS to force it to hand over client proxy voting data
The SEC filed a subpoena enforcement action on 4 September in the Eastern District of Pennsylvania after ISS refused to fully comply with an administrative subpoena issued on 21 July, part of an examination opened in March into whether its proxy recommendations meet its fiduciary duty. ISS argues the request raises First Amendment issues and could expose clients to retaliation. ISS and Glass Lewis together hold more than 90 percent of the proxy advisory market.
CNBC
CONVOGovernance detail worth carrying if you interview at an asset manager or a family office: two firms decide more than 90 percent of proxy advice, and the regulator now wants the voting data behind it. Ask what a firm does when ISS and its own analysts disagree, which is a question that lands.

Fitness & Performance

4 stories

Last 7 daysTwo usable synthesis papers, and they test the same question from opposite ends. One asks what a supplement adds on top of strength training in older people with sarcopenia and finds the training carries almost all of it. The other asks what a rubber band does for people with type 2 diabetes and no gym, and finds moderate certainty gains in blood sugar and grip. The rest of the seven days was mechanism work, animal work, paywalled trade pieces or repeats.

This quarterThe quarter bucket has spent ninety days ranking add ons and finding small effects: whole body vibration and blood flow restriction on 29 June, creatine on 6 July, anti obesity drugs for maintenance on 9 July, photobiomodulation on 20 July, keto at matched energy on 13 August. Since late August the briefs moved onto the instruments, with smartwatch calorie error on 2 September, doubly labelled water reallocation on 4 September, and ReacStep on 5 September showing the trained proxy improve while falls did not. Today swings back to the input side and the old ceiling pattern holds, in the sharpest form yet: HMB, the supplement sold hardest as the muscle preserver, added nothing significant to grip once strength training was already in the plan.

The lead

Across seven trials, training plus protein moved grip strength and adding HMB did not

The authors are explicit that seven trials, small samples and moderate to high risk of bias make this promising rather than settled. What the numbers do support is an ordering: the training and the protein are doing the work, and the specialist supplement is not adding a measurable second layer.

A systematic review and meta analysis published 4 September pooled 7 RCTs and 472 older people with sarcopenia doing strength training with whey protein or HMB. Overall the combination raised fat free mass (ES 0.39), maximal handgrip strength (ES 1.38), chair stand (ES 0.33) and 6 minute walk (ES 0.33). In the subgroups, whey plus training improved grip (ES 0.49) while HMB plus training did not reach significance, and neither changed fat mass or knee extension strength.

Medicine (Baltimore), via PubMed

sharpensmuscle-is-capitalThe asset is built by the training and fed by the protein, and the supplement marketed hardest at protecting muscle added nothing significant once training was in place.

SHINOMEUse it to answer the supplement question in a client consultation without a debate: the pooled evidence puts strength training first, protein second, and shows the branded muscle supplement adding nothing measurable to grip on top of those. Saves clients money and keeps the plan to two decisions.

1
Fourteen trials say an elastic band at home moves HbA1c and grip strength in type 2 diabetes
A meta analysis of 14 RCTs and 684 adults with type 2 diabetes found elastic band resistance training cut HbA1c by 0.41 and fasting plasma glucose by 0.79, and raised handgrip strength by 2.67 kg, all at moderate certainty with no heterogeneity on grip. Muscle mass favoured the bands but at low certainty with a confidence interval close to zero, and walking and sit to stand results were too heterogeneous to trust. Published 21 August and indexed this week, and adverse event reporting was too incomplete to judge safety.
Frontiers in Public Health, via PubMed
supportssimple-systemThe cheapest, most portable kit produced the moderate certainty results, so removing friction did not cost the outcome.
SHINOMEGives you a defensible entry option for Shinome clients who will not join a gym or travel with one: bands are not the compromise version on the outcomes that were measured well. Keep the claim to blood sugar and grip, since the muscle mass and walking numbers do not hold up.
2
A study of 682 older adults found no link between the ACTN3 gene and sarcopenia, and one exploratory activity threshold
In 682 community dwelling adults aged 65 and over examined between 2021 and 2024, the ACTN3 R577X genotype showed no significant association with any sarcopenia component. In the subgroup of women carrying the XX genotype, activity above 441 MET minutes a week was associated with worse five time chair stand performance, a U shape the authors call an exploratory finding needing external validation. The design is cross sectional and activity was self reported by questionnaire.
Scientific Reports, via PubMed
sharpenswalk-firstNothing here changes the sequencing: the genotype test was null, and the 441 MET minute figure is exploratory, self reported and cross sectional, so it cannot set a dose.
SHINOMEKeep the null, not the threshold, if a client arrives with a DNA fitness report: the gene test predicted nothing here, and the one number people will quote comes from a subgroup of a subgroup in a cross sectional study. It is a clean example of how to read a headline before changing anything in a plan.
3
The week's other research was mechanism work, animal work or already run
The three resistance training RCTs PubMed indexed this week were all already covered: the isometric corticospinal trial on 4 September, the 12 week supervised training trial on 4 September, and the blood flow restriction ACL trial, dropped for the sixth time because its epub date is 8 January 2026. The tyrosine and lifespan story doing the rounds is a 2025 Aging-US paper whose authors state plainly that it did not test supplements, and the blueberry compound story is cultured mouse muscle cells.
Shinome Brief
SHINOMETwo things to not act on this week: the brain supplement lifespan headline measured blood levels in the UK Biobank and not supplement use, and the blueberry fat burning story is mouse cells. Both will reach clients, so have the one line ready rather than the study.