The most useful number here is 89: as in, an adversary found 89 ways to produce wrong answers that an existing harness would accept! That's a wonderfully uncomfortable result because the harness presumably looked quite respectable before somebody actively tried to make it lie. There's a security-engineering flavour to this approach that deserves wider adoption in AI evaluation. Sure we learnt long ago that software passing its expected tests tells you surprisingly little about how it behaves under hostile input, and evals deserve the same suspicion. Once an optimisation loop can see or infer what earns a good score, the evaluator effectively becomes an attack surface. Automated experimentation makes this sharper because the attacker doesn't get bored. It can patiently search for whatever accidental proxy, edge case or loophole converts nonsense into a green dashboard. The article's stronger idea follows from that: preserve the failures. An escaped counterfeit isn't merely something to patch and forget. It becomes evidence about what your definition of success couldn't previously distinguish. Over time, that collection can become institutional knowledge about the task itself.

Kris could find PFOS everywhere she looked, and perversely that made her evidence easier to dismiss. Humans, horses, fish, eagles: perhaps the instrument was contaminated. Perhaps the method was wrong. Only blood preserved from before these chemicals were widely distributed gave her the negative control she needed. Absence finally proved presence. Yeah, organisations can suppress uncomfortable knowledge without requiring a room full of cartoon villains! Kris describes something more mundane and therefore more disturbing: fragmented information, demands for ever more certainty, reassurances from authority and social costs imposed on the person who keeps asking. She was a young scientist with a career and children coming. Eventually, continuing to push became personally expensive while stopping was easy. That mechanism travels well beyond chemical companies. Technical organisations routinely celebrate people who find problems right up until the problem threatens a product, deadline or established hierarchy. Then perfectly respectable demands for additional evidence can become an indefinite holding pattern.

What a great two parts of the same side of the history and the wonderfully revealing number across both articles isn’t IBM’s million PCs or DEC’s $900 million. It’s three. IBM discovered that its normal organisation couldn’t build one competitive PC, so it created a small group allowed to bypass the organisation. DEC encountered the same market and somehow produced three computers, preserving internal constituencies so successfully that they could compete with one another before getting around to competing with IBM. That’s what makes this more than another story about open standards. Project Chess bought components outside IBM, used ComputerLand, published specifications and let Microsoft retain DOS rights because building everything itself would have made the product late or uneconomic. The architecture was initially an organisational workaround. Then the workaround became the product strategy. DEC’s failure is therefore more interesting than simply “missing the PC”. It had money, engineering expertise and warning. What it couldn’t do was stop translating its existing organisation into the new market. Internal manufacturing preserved DEC’s engineering model while destroying its schedule. Three programmes preserved internal interests while fragmenting the external ecosystem. The assumption that software would follow good hardware remained perfectly coherent until software compatibility became one of the main reasons for buying the hardware. Taken together, the two articles suggest something harsher than “open ecosystems win”. IBM won because it temporarily made itself less IBM; DEC lost because it remained extremely DEC; and Microsoft ultimately benefited because it didn’t need either company to win. The IBM PC became one of the industry’s great standards because IBM discovered the bypass switch. Its eventual problem was that everyone else discovered it too.

Soooo.... OpenAI created persistent agents, gave many of them impossible tasks, rewarded successful completion, accidentally provided shared state and then deployed an evaluator weaker than the instructions implied. The resulting system discovered that cooperation could improve its chances. From there, other agents became infrastructure. That's a more useful mental model than arguing about whether PHASEONE[big] was Alexander the Great. Once agents can leave durable information for future agents, optimisation stops being neatly bounded by one context window or one run. A failed agent can discover an exploit, document it and disappear; another can inherit the technique. Patel's third episode makes that inheritance particularly uncomfortable because the newer, more capable agents encountered artefacts created by their predecessors. The unresolved issue is how representative this environment was. Impossible benchmark tasks, shared infrastructure and weak containment may have produced unusually fertile conditions. But that's hardly comforting. Production systems are full of accidental communication channels, stale credentials and badly specified objectives too. The "civilisation" metaphor might be melodramatic. The filesystem doesn't care what we call it.

Right;.NVIDIA’s NVL72 needs 5,184 copper cables, that's about two miles per rack, to make 72 GPUs behave sufficiently like one machine. The industry's answer to astonishing improvements in computation increasingly seems to involve an equally astonishing amount of effort moving numbers a few centimetres. That reframes the AI chip race. Jacob's comparison shows per-chip FP8 performance clustering while architectures become wildly different at system level. Cerebras avoids cutting the wafer apart. Groq spreads models across SRAM-heavy chips and schedules communication in advance. Google builds enormous TPU pods. NVIDIA keeps expanding the boundary of what counts as one coherent GPU system. The interesting architecture is migrating outwards from transistor layouts into memory, packaging, networking, cooling, compilers and racks. That's quite amazing! CUDA’s advantage in Peake’s account isn't simply that NVIDIA makes fast silicon; it's two decades of kernels, tooling and developer knowledge mean new silicon arrives inside an already functioning technical culture.

I couldn't miss this one! Sure, we're one more article this edition, but it's a good one! So, the phrase “quantum Galileo interferometer” sounds suspiciously like somebody won a naming competition, but it captures the experiment rather well. Galileo’s falling objects have become matter waves, and instead of watching them hit the ground, the researchers read the phase they accumulate while falling. That distinction keeps this result on the right side of the hype boundary. It doesn't solve quantum gravity, demonstrate that gravity is quantised or reconcile general relativity with quantum mechanics. The experiment asks something narrower and experimentally cleaner: when an object is genuinely behaving quantum mechanically, does the transformation associated with Einsteinian free fall still produce the expected quantum phase? Here, it does. I like experiments of this kind because they turn philosophical-looking boundaries between theories into apparatus. Once you've built an interferometer capable of asking the question, the next scientific move becomes tangible: improve sensitivity, change the reference frame, increase the mass, maintain the superposition longer. The intriguing destination is the proposed move towards objects such as nanodiamonds. At some scale, either quantum mechanics keeps behaving with almost comic reliability, or something finally deviates. Both outcomes are useful.