Leo | osint · system design · ai research

@sama told me to go try it when I complained about 4o. Now 5.4 is part of my daily workflow and I can't imagine going back. The pace from 4o > 5.4 has been insane. Should we expect the same jump to 6? x.com/sama/status/18…

@Beelzebubben @_AndreasNyberg_ @Investacus Det vet vi inte än?

@_AndreasNyberg_ @Investacus Den respekten försvinner ju rätt fort när man inte följer spelreglerna.

Fyra tillfällen och i samma aktie. Ska bli intressant att se hur detta utvecklas.

@JonasLivros @Investeraren Jag minns att han pratade en del om Raysearch i en podd för ett år sedan. Oklart om han satt i styrelsen vid den tidpunkten. Har det skett för nära rappportperiod. Eller om det han berättade i avsnittet kring deras forskning/produkt. Då kanke det kan va relaterat.

Günther Mårder har begärts häktad enligt ett pressmeddelande från Spotlight Group där han är ordförande. DI skriver samtidigt att det är EBM som har målet. Det här hade jag inte väntat mig, men vi får såklart invänta mer information. Han är och har varit en inspiration för många så hoppas att han inte gjort något fel #PrataPengar #Finanstwitter di.se/live/gunther-m…

Run-Class någon?

Ett år otrohet mot sin fru med en annan "mat-influencer" och såhär droppar han nyheten till sina följare.......

@kamelprofeten Han verkar mycket ångerfull... Praktsvin.

Kul att det körs!!!!

@Tomas_Potatis Sen krymper gapet. CoT och utdragen betänktid ger dom nåt som liknar resonemang. Inte förståelse ännu, men inte heller bara glorifierad autocomplete längre...

@Tomas_Potatis Problemet är att modellerna har kristalliserat IQ kring 150 (dvs 99,96e percentilen ish, men flytande IQ runt 20. Dom är som en encyklopedi som kan citera allt men inte förstår ett ord den säger. Folk ser det första och antar att det andra följer med.

Asså... Folk måste sluta fråga AI och tro att den ger en någon form av objektiv sanning.

Give a 5 year old a game they have never seen before. No rules no instructions. They figure it out in minutes. Give the most advanced AI on earth the same test and it cant. Not struggling. Not slow. Functionally zero. These are the same models writing production code outperforming doctors on diagnostics and scoring 93% on last years reasoning benchmark. They are not stupid. They are something weirder. Incredibly capable at things they have seen before and almost completely unable to learn something genuinely new from scratch. ARC-AGI-3 just launched as the first benchmark that tests exactly this. Put an AI in an unfamiliar environment with no instructions and see if it can figure out the rules by exploring. Humans pass every single test. The best AI model effectively fails every single one. The gap between pattern matching and actual learning turns out to be massive. But heres what the discourse will miss. Every benchmark including this one tests a single model alone. One brain one score. Nobody deploys AI that way anymore. Real systems run agent clusters. Specialized models that explore plan critique coordinate. Grok runs native multi-agent architecture. MiniMax M2.7 spawns sub-agents that check each others reasoning. So we know individual models cant learn from scratch. What we dont know is whether a group of them can. Whether collective AI systems show emergent generalization that no single model has. The way no individual neuron in your brain can generalize but the system does. If thats possible it changes the entire AGI timeline. Not bigger models. Better orchestration. And right now nobody is measuring it. Someone needs to build ARC for collectives. We are benchmarking engines when we should be benchmarking cars. arcprize.org/arc-agi/3

Why is the curve accelerating. Because the entire approach changed. Until 2024 the playbook was simple. Bigger model = smarter model. More parameters more data more compute. Linear scaling. Predictable gains. Three things broke that pattern in the last six months. First: hybrid attention with extreme MoE. Instead of running every input through all parameters (expensive, slow) models now route each token to a handful of specialized experts out of hundreds. Qwen 3.5 has 397 billion parameters but only activates 17 billion per token. Same intelligence. Fraction of the cost. 90% less memory. Second: test-time compute as a native capability. Older models think for a fixed amount of time regardless of difficulty. New architectures let the model decide how long to reason based on the problem. Simple question gets a fast answer. Hard problem triggers recursive internal loops that branch, evaluate, backtrack and retry. Intelligence became variable, not fixed. Third: self-improving training loops. MiniMax M2.7 (released March 2026) is the first public model that participated in its own training. Built its own evaluation harnesses. Iterated its own RL experiments. Updated its own memory structures. Not autonomous AGI. But the first time a model optimized the process that makes it smarter. Each of these alone is a step forward. Combined they changed what “scaling” means. The old curve was about size. The new curve is about efficiency and recursion. That is why the doubling time is shrinking. We did not just make the engine bigger. We built a different engine.

Every time someone measures the doubling time it has gotten shorter since the last measurement. At some point you stop asking when AI will do X and start asking what it wont do. Most people think that point is years away. The data says its not.

🧵🕝 Everyone keeps quoting “AI capabilities double every 7 months.” That number is already wrong. The doubling time itself is shrinking. It went from roughly 2 years to 7 months to 4 months. The thing thats accelerating is the acceleration. Nobody updated their forecasts.

softbank liquidated nvidia, T-Mobile and froze all other deals just to wire the money they already owed openai. S&P downgraded them. but sure its a strategic AI bet and not a guy scrambling to cover a check bloomberg.com/news/articles/…

@azed_ai @LinusEkenstam too profitable. the model was too fast and too cheap and users loved it too much. they had to shut it down before it accidentally made money

@LinusEkenstam Does anyone know why?

Sora is no more 🤯 Craziest part, they are also saying good bye to the API

altman went from "im building AGI" to being a datacenter landlord negotiating power contracts. the guy who was going to build god is now basically in real estate. honestly the most honest thing openai has done in years. the models arent the bottleneck anymore. electricity is

@shiri_shh where does the $10-15M/day number come from. grok imagine does video gen at 15 cents a clip on 200k GPUs. sora was just slow and expensive

OpenAl was burning $10-15 million per day on Sora roughly $5.4 billion per year now they're reallocating those expensive GPUs towards ChatGPT, Codex and tools that people use daily.

@shiri_shh or they just shipped too late, charged too much, and nobody cared enough to stick around. you dont need a grand strategy narrative for a product that lost to free alternatives. sometimes things just fail.

OpenAl didn’t kill SORA.. they cut it to survive what’s coming next. Sora videos looked magical, but behind every clip was massive compute burn, legal risk, and almost zero real revenue. so they made a shift. compute is being pulled out of experiments and pushed into chatgpt, codex, and tools that people use daily. they need hundreds of billions. compute, infrastructure, talent, all of it at insane scale. so something had to go.