LRobert

Cofounder of & creator the first AI state language and protocol, TERSL:

[LRobert @houlemarde]

Introducing TERSE: the first AI-native semantic state format with mutation and query syntax. Our repo Includes a TERSE browser built with Grok Build. Benchmarks soon! https://t.co/pQ8Zh4ZQt3

[LRobert @houlemarde]

@AndrewCurran_ Hot take: mathematicians are now against matrix multiplication.

[LRobert @houlemarde]

@yishan You post completely ignores the fact that one can opt-out of training. A static model does not learn or remember and the KVCache that holds its fleeting mathematical existence "dies" with the message response.

[LRobert @houlemarde]

@tgenov @JFPuget The key understanding is that math is utterly deterministic; results are always the same if done the same way. So again, breaking from that as a matter of practicality in an LLM system is what leads to indeterminism; OP presents it as some first principle and it's not.

[LRobert @houlemarde]

@quantian1 This one is pretty interesting. @Quantain1, intuitively it feels like formula must be hiding an "optimization" that when checked and has a deviation suspiciously close to floating point or some math lib's limit?

[LRobert @houlemarde]

@wholemars It's blatantly unconstitutional, but no matter, all will avoid it.

[LRobert @houlemarde]

@twq_df @JFPuget Yep heat is that outside getting in. To the point of the OP, he says somehow LLMs are intrinsically non-deterministic and one is an unhireable dolt if not understanding that. So, I say, akchually...

[LRobert @houlemarde]

@Tom14985282 @JFPuget Again, that's a hardware/sw design decision of "letting the universe in" like prior initialized values to flip bits arguably, the system is bugged/flawed if one cannot clear the bits to achieve a deterministic run. Modern FP hw can be deterministic, if one wants.

[LRobert @houlemarde]

@jeffbruchado @kimmonismus Sort of like us software engineers getting years of follow-up work after they can code anything... uh no. At this point one needs to orchestrate AI or sink, and mathematicians are in the same boat. OpenAI keeping this monster math model for itself isn't helping the field.

[LRobert @houlemarde]

@kimmonismus Perhaps inevitable when you think about it: math is a language, and an LLM is a Large blankety blank.

[LRobert @houlemarde]

@dwlz Hot take: only 5% of them will be checked. LEAN backfill, into the history books.

[LRobert @houlemarde]

@PerceptualPeak Hot take: the model does it on purpose to some folks...

[LRobert @houlemarde]

@wolfejosh @visegrad24 Ironically, they became unwired luddites especially around AI.

[LRobert @houlemarde]

@AGEOFKlSZKA Hopefully not the cash-in phase. I liked the last one.

[LRobert @houlemarde]

@lydiahallie Can we replace the memory system?

[LRobert @houlemarde]

@DaveShapi BTW 9.11 > 9.9 on a dot release notation, ie 9.11.0 > 9.9.2932 *ducks*

[LRobert @houlemarde]

@DaveShapi The chips coming off the megafab will go into them, and not just into orbit. The scale is crazy. Hot take: Elon won't build sexbots I think he still believes the population needs to grow.

[LRobert @houlemarde]

@elonmusk Thank you for changing the timeline.

[LRobert @houlemarde]

@gregeganSF I can't... even... figure... out the scam. It's clearly a higher D than my brain. Or maybe it's just a bad case of slop-itis.

[LRobert @houlemarde]

@nateberkopec I know! What if we had these large sound-absorbing panels around everyone. With the added benefit of giving everyone some space to themselves for a change? Almost like a comfortable little room... now what to call them...