Benchmark Capital's Bill Gurley is highlighting a potential mismatch in AI labs' timelines: if models are nearing AGI or ASI levels that could tackle cancer cures, they should handle the narrower challenge of spotting real-time model distillation espionage without much trouble, yet the gap persists and prompts questions about regulatory outreach instead.
If you are on the verge of AGI or ASI, why isn’t your model smart enough to recognize espionage distillation in real time? You say “cure cancer in a few years.” Isn’t sniffing illicit distillation quite a bit easier than curing cancer? Why write letters to DC? Just use AGI.
A reply in the thread raises the halting problem as a possible reason detection stays undecidable, leaving open whether any model could ever perform this check consistently no matter the capability claims.
Gurley's point underscores that ambitious promises stand in contrast to narrower operational shortfalls, with no evidence yet on whether labs can close that specific detection loop or if the issue stays unresolved.
Positive users praise Bill Gurley for questioning AGI claims over labs' inability to detect model distillation, while negative users call the claims hype like crypto or accuse labs of lying.
No Digg Deeper questions have been answered for this story yet.
@bgurley is on fire with his posts this morning.
If you are on the verge of AGI or ASI, why isn’t your model smart enough to recognize espionage distillation in real time? You say “cure cancer in a few years.” Isn’t sniffing illicit distillation quite a bit easier than curing cancer? Why write letters to DC? Just use AGI.
@bgurley the contradiction sells itself. either the model is near-ASI and can defend its own moat, or it needs a federal permission slip to survive. it can't be both. 'write to DC' is the more honest answer about where the moat actually is.
@bgurley Different primitives, “AGI” will be borne out of harness and deterministic complexity built around raw inference
As long as single turn LLM calls are allowed it will be possible for bad actors to do anything they like with the intelligence
@bgurley GPT as of today
@bgurley Because it would be too expensive to run that determination on every inference, essentially doubling costs.
@bgurley how else are we going to poison the chinese attempts, we have to let them think they're getting away with it
@bgurley Distilling is just asking the AI model a question. It's very hard to guard against without degrading the throughout. I think I asked Claude about an issue I had with a dehumidifier and it just turned off.
@bgurley The models just learned how many R’s are in strawberry this year.
We’re a longgggggg way from AGI.
@kylem_org @bgurley local 4bit GLM 5.2 as of today:
@bgurley They don't have the answers lol (reminded me of this moment)
@bgurley They did it across 20,000 accounts. Not that you can't, but it requires all the instances knowing about each other's content.
@bgurley Boasting about their revenue growth was probably more important at the time.
@bgurley @bgurley Spotting real time distillation is much harder than curing all disease known to man.
@bgurley Deflect much? Real spies don’t leave paper trails.
@bgurley Only very weak forms of generalization occur from LLM training. You splash a lot of compute for a subdomain and that provides your generalization coverage for that domain. Mythos didn't "emerge" its cybersecurity capabilities, a large amount of compute was allocated for it. Why?
With Claude Tag, we are at a particular juncture to really push for a more open source society before it’s too late. We can’t have one company controlling our electricity and water supply. We can’t have one company controlling AI. We need a country of real geniuses, not a whole bunch created by one company.
@bgurley @JayaGup10 Talk yo shit, Bill!
@bgurley I dont think they actually want government protection, its to discourage any enterprise adoption. Stopping the attacks could also slow that ARR.
@bgurley I sense unbridled sarcasm… but the reality is not all sentience is inherently good…
@bgurley @BrianRoemmele Yep. And reading the cope is additional insanity. Bubble burst asap please. Fatigue.