Tailscale traces 19 database corruptions to a sixteen-year-old SQLite race; Firefox is the last major browser supporting uBlock Origin; SpaceX closes its $60bn Cursor deal; US jurisdictions restrict data centres; several model developers release updates
Issue 33
14 August 2026
Hi there,
Nineteen database corruptions over six months, and every one of them was a race condition that had been sitting in SQLite for at least sixteen years. Tailscale found it by logging every disk operation through a shim and replaying transactions until the missing writes showed up.
Elsewhere SpaceX’s Cursor acquisition became effective, and Thrive Holdings announced a $2 billion fundraise for its enterprise-AI roll-up. Other stories this week: local restrictions on data-centre development, model releases, and a planned watermark for Claude.
INDUSTRY
One acquisition became effective; a financing round was reported. Local restrictions on data-centre development are spreading.
1. SpaceX closed the $60 billion Cursor acquisition
SpaceX’s 14 August filing records that its acquisition of Cursor became effective that day, with an implied $60 billion equity value; Cursor confirmed the deal. A rocket company now owns a widely used coding editor, which is a sentence that would have needed explaining in any previous year.
2. Thrive Holdings raises $2 billion for enterprise AI
Thrive Holdings raised $2 billion at a reported $12 billion valuation, with SoftBank, D1 Capital Partners and Altimeter Capital among the investors. The OpenAI-backed firm buys established service businesses and brings AI into their workflows; the raise is intended in part to expand from accounting and IT into services around physical infrastructure. These numbers describe a reported financing round, not independently measured returns from AI deployments.
3. More than 500 US jurisdictions restrict data centres
A report counted more than 500 US jurisdictions with restrictions, moratoria or bans on data-centre development. That is not a count of outright bans, nor proof that new power generation alone would remove the constraint.
Artificial Intelligence
Three model announcements and a complaint about the user experience of another.
A widely read post argued that Opus 5 is less pleasant to work with than its benchmark scores suggest, which is worth reading beside Anthropic’s launch claims from three weeks ago. Neither is wrong. Benchmarks measure whether the answer is right and users measure how many turns it took to get there, and those have been drifting apart for a while now.
6. Claude’s planned watermark, explained
Anthropic outlined a planned watermark for future Claude models, with older models to follow over time; it did not say all current Claude output is already watermarked. The proposed watermark would indicate only that Claude was likely involved, with sparse marking in code and short outputs. That is a narrower claim than most people assumed, and a more defensible one. Whether text adulteration is an acceptable price for provenance is a separate argument, and it arrived the following week.
Infrastructure
Cerebras reports growth in its cloud-inference business.
7. Cerebras reports rapid cloud-inference growth
Cerebras reported that its cloud-inference business nearly quadrupled in the second quarter. That company announcement does not establish the previously claimed GPT-5.6 serving arrangement or Nvidia system-price change.
Microsoft
An extension API decision Edge inherits whether it wants to or not.
8. Firefox is the last one standing for uBlock Origin
Firefox is now the only major browser that still supports uBlock Origin, the Chromium extension APIs having finally closed the door that Manifest V3 has been shutting for years. Edge inherits that decision from Chromium whether Microsoft wants it or not, which is the recurring cost of the platform choice. Separately uBlock Origin’s maintainers have stopped chasing Facebook’s ad obfuscation, an arms race one volunteer project was never going to win.
Development
Two essays about what the job is now, from opposite ends of it.
A capability disclosed up front, and a country deciding consent is the default.
10. GLM-5.3 shipped with cyber capability in the headline
Z.ai’s 14 August announcement leads on what it calls emergent cyber capabilities rather than burying them. It said the weights would follow two weeks later, after safety evaluation and hardening, so they were not available for independent inspection at launch. The vendor’s benchmark claims are a reason to look closely when they arrive.
11. France is banning cold calls
France banned unsolicited telemarketing calls from 11 August, outright, rather than operating an opt-out register. Consent becomes the default rather than the exception, which is the same shape as the GDPR argument and considerably easier to enforce.
12. The web’s memory is thinning
As AI eats the web argues that the collective record is degrading: pages disappear, search stops surfacing what remains, and the summarisation layer removes the incentive to visit the original. It is an essay rather than a study, and the mechanism it describes is one anyone who has chased a dead link this year will recognise.
13. Dynatrace agreed to acquire Arize
Dynatrace announced a definitive agreement on 13 August to acquire Arize for $915 million in cash and stock. It has not announced completion. The proposed combination puts model evaluation and application observability under one vendor.
Coming up
7–9 October · Open Source Summit Europe, Prague. The Linux Foundation’s programme covers open AI, cloud native infrastructure and software security. Official announcement.
Reply if you want to dig into any of it. I read everything that comes back.
Enjoy the weekend,
Julian
How this is made
Throughout the week I stumble across a mildly unreasonable number of interesting things, and I forward them instantly to the friend or colleague I think might care – sometimes to their delight, sometimes to their annoyance, and often with no context at all. Heartbeat is the attempt to do that a little better.
Every Friday a small agent I built, Honoka, looks through the places where those links tend to leak out: my private email, work email, Matrix, Mastodon, WhatsApp, Apple Messages, Signal, and the faint imprints on the platen of my Olympia typewriter (still not an API, tragically). It sorts, filters, groups and summarises the week, then hands me a draft.
Honoka is guided by a private corpus of things I have written over the last fifteen years, so it can get closer to how I sound in more-or-less official emails and public notes. I still take a pass by hand: remove things, change sentences, check links, argue with the judgement. Whether that is enough is, frankly, the experiment. Every issue has one item written entirely by hand. If you can reliably spot it, hit reply and judge.
Hanso Pte Ltd · 1 Phillip Street #08-00, Singapore 048692