Overview
Today’s posts circle around a practical question: what counts as proof that AI is doing useful work? Microsoft’s leaders argue for containment and independent checks, a cancer-cure thought experiment tests the limits of explainability, and developers report faster software reverse engineering. Elsewhere, the discussion turns to computing costs, research training and the schools still missing basic internet access.
The big picture
Chamath Palihapitiya backs independent checks on advanced AI
@chamath backs Satya Nadella’s argument that businesses should treat advanced AI as a potential insider risk, not simply a capable assistant. His central point is organisational: separate the systems doing the work from the people and processes that verify it, measure it and approve it.
Mustafa Suleyman makes a related case for containment, comparing the challenge with safety controls for aircraft, medicines and nuclear materials. Neither argument requires settling when superintelligence will arrive. Permissions, audit trails and independent checks are useful requirements for systems businesses deploy today.
Andrew Curran challenges the need to understand an AI-discovered cure
@AndrewCurran_ responds to a thought experiment from Terence Tao’s Caltech lecture: what if AI produced a cancer treatment with strong trial results, but nobody understood its mechanism? Curran argues that a patient with stage 4 cancer would care more about whether it worked.
The distinction worth keeping is between understanding a treatment and having evidence that it is safe and useful. Medicine can benefit from treatments before their mechanisms are fully understood. That does not make an AI model’s own assessment a substitute for independent clinical testing. The debate is about what evidence is sufficient, not simply whether a black box is acceptable.
NVIDIA reportedly explores a deal with Reflection AI
@zerohedge highlights Financial Times reporting that NVIDIA is in talks about acquiring Reflection AI or deepening its investment. The discussions are described as early, with an acqui-hire also reportedly under consideration. This is not an announced deal.
Reflection’s work on open-weight models gives the talks a broader context. NVIDIA already supplies much of the hardware used to build and run AI; a closer relationship with a model developer would extend its involvement in the software running on those chips. Open weights and concentrated corporate ownership can coexist.
Matt Pocock revisits the cost of fast, messy coding
@mattpocockuk quotes John Ousterhout’s description of the “tactical tornado”: a programmer who produces code quickly but leaves difficult maintenance work behind. Managers see rapid output; colleagues inherit the cleanup.
Replies connect that older warning to AI coding agents. Producing more code is not the same as improving a codebase. If teams reward visible speed without accounting for maintenance, an agent can multiply the same problem that already existed.
A.I.Warper reports rebuilding a Steam game with AI tools
@AIWarper says the Reverse Engineer Anything toolkit decompiled a purchased Steam game into C# and launched it in Unity in under 30 minutes, using Opus 5.5. Matthew Berman also shares a demonstration of the toolkit’s agent-led software analysis.
These are reported demonstrations, not proof that arbitrary software can be reconstructed with the same ease. Replies point out that established Unity tools already handle parts of this work. The interesting question is how much agents reduce the manual coordination between those tools. Buying a game also does not automatically grant permission to redistribute its code or assets.
Gregor Zunic questions the price of running models locally
@gregpr07 contrasts a $20,000 local machine running GLM 5.3 at 70 tokens per second with an overnight run through OpenRouter that cost him $10. For occasional use, that is a reasonable challenge to an expensive hardware purchase.
The replies make a different case: local computing buys control, privacy and access to model variants that hosted providers may not offer. A single night’s bill cannot settle the economics, but it does expose the question buyers need to answer: are they paying for cheaper computation, or for independence?
DogeDesigner highlights India’s school internet gap
@cb_doge cites the Education Ministry’s UDISE+ report for 2025-26, which records internet access in 67.4% of India’s schools. Of 1,466,682 schools nationwide, 477,491 were listed without it.
The post makes a case for Starlink in remote areas, but the figures establish the connectivity gap, not which service should fill it. Satellite access deserves consideration where terrestrial networks are difficult to reach. Cost, regulatory approval, reliable electricity and usable classroom equipment still need to be part of that discussion.
Geoffrey Hinton recommends a PhD for aspiring AI researchers
@MIT_CSAIL shares Geoffrey Hinton’s advice to computer science student Samuel Zhang: he would still pursue a PhD if his goal were to become a leading AI researcher.
That is a narrower claim than saying doctoral study is the best route into every technology job. Research training offers time to investigate difficult questions, learn from experienced researchers and develop original work. For someone aiming at research rather than product development, Hinton’s answer is a useful counterpoint to the pressure to skip formal study.
Jason Fried praises Hector Serrano’s account of building an ads programme
@jasonfried points to a detailed video from Hector Serrano, the 22-year-old head of content at WisprFlow, explaining how he built a performance-based creator advertising programme. Fried’s reaction is less about the job title than the evidence: here is someone explaining what they actually did.
Serrano describes starting with a small group of committed creators, providing coaching and rewards, then growing to 50 active creators. It brings the day’s discussion back to a plain standard: claims of capability are easier to assess when someone can show the work, explain the choices and account for the results.
