Damn! The number that sticks with me is 13 hours... (only!) Not because that's astonishingly fast in isolation, but because five unrelated devices fell apart under roughly two working days of agent attention. Reverse engineering used to have a useful economic defence: even mediocre firmware could be annoying enough to understand that attackers needed a reason to bother. Charz experiment suggests that friction is disappearing! The Key Light Mini is the sharpest example. Elgato actually signs its firmware, which sounds reassuring until an exposed memory-write path lets the running system disable the check. Security concentrated at one checkpoint becomes surprisingly fragile when an agent can patiently map everything around it. That changes the sensible default for endpoint design. A USB microphone may be a small computer with writable firmware, privileged interfaces and enough undocumented behaviour to deserve suspicion. We have spent decades making computers distrust files from the internet while cheerfully plugging tiny mystery computers into them. That asymmetry is starting to look expensive.

The oddest detail here is 28x28km; as cyclone intensity is driven by fine-scale processes around the storm’s core, so conventional wisdom has pushed forecasting towards increasingly fine spatial resolution. WeatherNext Cyclones apparently gets competitive intensity forecasts from inputs roughly 100 times coarser than traditional models. Google admits its scientists don’t fully understand why. That deserves more attention than the "decade of progress" headline. If the result holds up, AI weather forecasting may be learning useful large-scale signatures that act as proxies for physical processes it cannot explicitly resolve. That's powerful, but scientifically awkward. A model can be operationally valuable while leaving meteorologists with a new problem: distinguishing genuine learned structure from correlations that happen to survive the evaluation set. The 1,000-member ensemble makes this particularly interesting. Cheap inference changes forecasting from producing a better single answer towards cheaply exploring a distribution of plausible futures. For emergency planning, the nasty tail of that distribution can matter more than its centre. So, open sourcing the weights and code is therefore the consequential next experiment. Independent meteorologists can test unusual storms, different ocean basins and failure modes rather than admiring Google's aggregate scores. If a coarse model keeps beating expensive high-resolution approaches outside Google's evaluation, meteorology may have acquired both a remarkably useful forecasting tool and a rather good scientific mystery.

Yeah, Chris tried the cheap, obvious automation first: run m2c across 1,830 unmatched functions and automatically accept exact matches. It found 17. The big productivity gain came elsewhere, from building an environment in which agents could repeatedly attack messy problems, retain discoveries and receive increasingly useful feedback. That distinction feels important for anyone deploying coding agents. The impressive part isn't that an LLM can emit plausible C from MIPS assembly. He built a small production system around the models: four parallel worktrees, explicit deadlines, shared compiler learnings, similarity search across workers and tooling that explains why two compiled functions differ. The agents become more useful because the surrounding machinery converts vague failure into actionable information. There's also a nice corrective to the usual "AI replaced months of engineering" story. Chris had already spent nearly two years learning this peculiar craft, the community had reverse-engineered parts of a long-dead SGI compiler, and specialist humans apparently rescued matches that might otherwise have stopped the project around 89 to 90%. The agents arrived on top of accumulated human knowledge rather than replacing it.

When I opened tmp.0ut, I immediately felt that little jolt of recognition you get when a screen looks like somewhere you used to go rather than something you're supposed to consume. If you spent years in terminals, dialling into BBSs, reading textfiles or admiring absurdly elaborate ANSI screens, the interface isn't decoration. It's part of the content. That's why the Doug McIlroy interview belongs here so perfectly. You're reading the inventor of Unix pipes talking about machines with tiny memories, 1200 baud feeling fast, debugging with printf, and the evening when pipes suddenly worked. Around him are articles about 57-byte ELF files, metamorphic code, binary golf and strange loader tricks. Different decades, same impulse: understand the machine deeply enough that its supposed boundaries become suggestions. Modern computing has made almost everything prettier and much less visible. tmp.0ut goes in the opposite direction. The characters are the graphics. Handles matter more than job titles. There is a mixtape. People spend ridiculous amounts of effort shaving bytes or abusing executable formats because it amuses them. For anyone who remembers BBS culture, that feels less like retro cosplay than continuity. And I think that's the nostalgia worth preserving. Not beige cases, CRT filters or simulated modem noises, but the feeling that a computer is a place you can explore. You type something, something strange happens, and curiosity tells you to keep going.

I think there’s a wonderfully awkward contradiction at the centre of this design; Samsung has gone to considerable lengths to make PIM look like ordinary LPDDR5X, then given ordinary memory operations decidedly extraordinary side effects. A read might perform a computation. A write might alter PIM state. Suddenly all the clever tricks processors have accumulated around the assumption that reading memory is harmless become liabilities. That makes the 614 GB/s figure almost secondary. Modern CPUs aren't fast simply because DRAM is fast. They’re fast because decades of architecture have been built around hiding the fact that DRAM isn't: caches, speculation, prefetching, out-of-order execution and interleaved channels. PIM gets closer to the data, then potentially asks software to surrender several of those mechanisms to use it safely. PIM itself is an old and seductive idea: if moving data costs too much, move computation to the data. Samsung demonstrates that the arithmetic is perhaps the easy bit. The difficult part is persuading fifty years of computer architecture that memory has quietly started thinking.