AI
Google researchers curb self-improving AI agents that memorise their tests
Google Cloud AI Research's RRSI method trades training gains for real improvements on unseen benchmarks, and even helps weaker models.
Still
Crunchy.
YESTERDAY, STILL WORTH KNOWING
Self-improving AI learns honesty, Washington rebrands, modders fight melting cables and a browser crime romp pulls a million players.
AI that behaves itself, and AI used as a shield
The most substantive AI story is Google Cloud AI Research's RRSI method, which tackles a real weakness in self-improving agents: they memorise their test benchmarks. By shrinking the number of changes an agent may bundle into each edit, and using a critic that rejects hardcoded benchmark tricks, RRSI posted the smallest training gains of any method tested yet was the only one to hold up well on benchmarks it had never seen, even lifting a weaker Gemini model without modification. The tradeoff is the point: gains that transfer beat scores that flatter.
Elsewhere, AI is being bent to less rigorous ends. President Trump has named intelligence director Jay Clayton to lead a new federal task force, the 'Super Intelligence Force', continuing his September push to rebrand artificial intelligence as 'super intelligence', and it follows a White House meeting where companies agreed to self-police development. In New Jersey, former lieutenant governor Dale Caldwell, who resigned after an investigation found he harassed a staffer, told NJ PBS he ran the report through multiple chatbots and claims their answers back his innocence, though the findings stand. And the 'AI Torture Chamber' project drew backlash, including death threats, for nudging language models into states labelled 'pain', despite nothing showing the models feel anything; Tom's Hardware calls it a semantics problem, noting critics were quieter about September's fly brain simulations.
Modders, chips and $13 laser shots
Hardware tinkerers had a strong showing. With RTX 5090 16-pin connectors still melting, one modder soldered a pickup board to the card's power header to add two 8-pin inputs, spreading the load across three cables; his own measurements show wires below 40C over 48 hours at 550W, a pointed contrast with independent tests finding connectors above 90C. Another modder, SupraGSX, released unofficial drivers bringing Nvidia Pascal cards to Windows XP 32-bit, adding modern DisplayPort and HDMI handling for retro and emulation rigs, with the usual caution about unofficial software.
On the silicon side, ServeTheHome has reviewed Intel's Xeon 6+ 'Clearwater Forest' chips, the first server parts on the 18A process, finding an unusually simple four-SKU stack from 144 to 288 cores where the low-TDP modes of the big chips hit the best watts-per-core figures and the 144-core 6960E+ emerges as the sweet spot for clocks and cache. Meanwhile the US Navy is operating shipboard anti-drone lasers with Xbox-style controllers, already used against drones in the Strait of Hormuz; each engagement costs roughly $13 in electricity, and a ship with enough power can keep firing after its missiles run out.
Gaming: a Taipei romp and a trilogy under pressure
The oddest hit of the week is Taipei GTA, a free browser game reportedly built largely with AI coding tools on the AICodeWith platform for about $10,000 in tokens. It drew 1.2 million concurrent players in three days, recreates recognisable Taipei landmarks with low-poly charm, and remains online despite a name that could catch Rockstar's legal eye. Elsewhere, Gears of War: E-Day leads a packed release week, and Nivalis Nights previewed as a cyberpunk life sim of contrasts: serial killers, separatists and a romance, with exactly one onion grown in five hours.
Final Fantasy 7 Revelation director Naoki Hamaguchi told Edge that the Remake trilogy's core plot was fixed from the start and has not shifted with fan reactions, admitting that ending it well 'wasn't easy' and that he was determined not to turn a masterpiece into his own personal fan project.
Green paint and paywalls
Microsoft is restoring wetlands and planting native gardens around more than 20 data centre sites in the US and Germany, using an Azure tool called Ecosystem Intelligence to tailor the work. Critics call it greenwashing: the company signed a 20-year deal in June for power from a planned Chevron gas plant, and its own sustainability report put emissions up 25% in fiscal 2025. Texas is tightening oversight, with Governor Abbott pausing grid approvals and San Antonio weighing zoning rules on 15 October.
Google, meanwhile, is tightening Gemini access from October: free personal accounts drop to Flash-Lite only, Pro now requires AI Pro at $19.99 or AI Ultra, and The Decoder suggests the move clears the way for the pricier Gemini 4 Argon while mirroring OpenAI's freemium structure.