Former OpenAI CTO does what Altman won't, releases a frontier AI model that's actually open

If you’re in the market for a frontier-class open weights model, your options are few and far between outside of the Chinese model houses. With the Wednesday release of a new model code-named "Inkling", an outfit called Thinking Machines Lab aims to change that. Founded in early 2025 by former OpenAI CTO Mira Murati, Thinking …

[Fido]

Thinking Machines was not a fictional company.

https://en.wikipedia.org/wiki/Thinking_Machines_Corporation

[that one in the corner]

The youngsters these days, they don't know their history, aren't making the Connections*...

* and probably don't even know of James Burke; that single-take walk down the Saturn 5 is still unbeaten for timing.

[Bebu sa Ware]

"not a fictional company."

I also recalled them as an '80s AI and supercomputing company.

It's surprising that some AI wannabe has not adopted the Sirius Cybernetics Corporation^† out of the orphanage of fiction — the tech bros have never shown any lack of chutzpah.

_{† or Cybus Industries.}

[Irongut]

Re: "not a fictional company."

As a spin off from Palantir most likely.

[T. F. M. Reader]

If Jurassic Park's Trivia section on IMDB can be believed an actual CM-5 by Thinking Machines Corp is shown in the movie. The Wikipedia page (cf. the original post) says what was shown in the movie was non-functioning dummies. The IMDB quotes Daan Sandee of Thinking Machines confirming it was a real computer. To quote the IMDB site (not Daan Sandee directly), "[T]he CM-5 super computer, used in the control room, was one of only two ever built to that size (1024 nodes). The other machine was at Los Alamos. The machine used in the movie, was sold in smaller segments after the scenes were complete. Mirrors were used to make it seem like more CM-5s were present."

Maybe "non-functioning dummies" means the CM-5 was not actually functioningduring filming of the movie? Regardless, "non-functional" != "fictional", and "fictional supercomputer maker" is certainly unwarranted.

[doublelayer]

It seems weird to use functional computers as props when empty cases, at most with something simple simulating the blinkenlights, would do just as well, but I suppose that's not sufficient evidence that they didn't.

[Pascal Monett]

975 billion parameters

So, when will get a billion parameter hallucinating fuckup ?

[Pascal Monett]

Grr. I should have wrote a trillion.

[Dan 55]

You're hallucinating wrong. You have to confidently declare you're right and the meatbags will fall into line.

[Bebu sa Ware]

A trillion parameters

975 billion parameters requires 8 x Nvidia's B300 accelerators each of which consumes 14.5 kW according to NVidia's specs.

A trillion I am guessing not a lot more but a quadrillion (10¹⁵) parameters might require 1000 x 8 x B300s if it scales linearly but likely a lot more. So 116 MW. ;)

Fairly sobering when one considers that the human brain requires in the order of 20W and a first class brain could entertain "imagine something unimaginable" without cooking itself which a quadrillion parameter model might do but is more likely reply "sorry Dave, I'm afraid I can't do that."

[Long John Silver]

More, more, more - bigger, bigger, bigger - BETTER? and for whom?

An initial 975 billion parameters for assignation of values during 'training' (i.e. exposure to whatever digital materials are to hand) represents an immense number of 'degrees of freedom' for model fitting. Given that a modest computing device with VRAM in the range 16-32 GB is capable of doing tasks (e.g. 'imaginative' text and image processing) not so long ago in the realm of fantasy, one wonders what this two terabyte leviathan can accomplish. Similarly, where do even larger models, not in the public domain, and never likely to be, lead?

Setting aside nonsenses to the effect that ever greater 'degrees of freedom' must lead to 'emergent properties' and a 'singularity', one wonders what mighty existential questions the technicians behind AI, and their supposedly prescient money-masters, have in mind. Cynical consideration of each of the mindsets leads to 'suck it and see' rather than the foresight of Asimov's 'Foundation series' books character Hari Seldon.

These considerations suggest requirement for urgent strategic thinking over the best uses of resources for large-scale computing. Two choices present: general purpose 'supercomputers' programmed by people for specific tasks (e.g. modelling solar activity), or differently organised computation centred upon immense 'black box' statistical models. The latter, have various kinds of proven utility, but reliance upon their cognition may pose an epistemological trap.

Unfortunately, a considered view of the wisest among mankind shall never emerge, nor ever be sought. Man's governance rests in the hands of a shallow 'political class' directed by minds obsessed with finance, personal power and dynastic enrichment.

[elDog]

Re: More, more, more - bigger, bigger, bigger - BETTER? and for whom?

You make my brain hurt. More.

[theOtherJT]

"If you prefer to evaluate the model on your own hardware..."

...on my own hardware?

Weighing in at 975 billion parameters, the model requires more than two terabytes of GPU memory

I don't think many of us have that sort of compute just lying around...

[Anonymous Coward]

Re: "If you prefer to evaluate the model on your own hardware..."

Frontier open-weight models are for corporations building anything intended to survive the AI bubble as well as the next update to ChatGPT and Claude.