spindump8930

That's overly generalized. Perhaps they became a buyer after google stripped top staff, but before that they were certainly a producer, and had many advances in cutting edge inference at the time.

"Commits are to private repos" why is this unusual? Very expected if OpenAI has private repos in a github org.

Words are meant to convey an idea to your audience! In some sense there's no "should" - either a word works or it doesn't.

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

They hired many accomplished researchers, with track records of both papers and work at other frontier companies. They gave those folks huge autonomy, huge budget, and 18 months - some would say at the expense of the rest of the company. Check out blind references to TBD. I'd call the strategy ruthless, but it certainly was effective.

Strong agree, a good way forward is letting user agent once again truly mean an agent on behalf of the user. And speaking anecdotally, I've seen different traffic "shapes" under the same UA, which leads me to suspect that some actors might be disguising traffic as more mainstream labs.

Well, potentially very little! It could be subscription usage, or just really inefficient code authored by an agent. It probably costs LESS in tokens to write a poor quality cloner that breaks upstream things.

The problem (and contrast with other approaches) is that mat muls requires synchronization. Arranging your networking and training structure to maximize compute and minimize communication is the main craft of ML training infra folks. In your example, yes you can compute layers on different machines (i.e. Tensor Parallelism), but you must be very careful in how you arrange it.

GPU failures are frequent enough that at a certain scale, you constantly have workers dropping out. Designing systems that can still keep training is very interesting!

All of those are good references. Other folks in the thread are missing distinctions between pre training (~the internet + curated sources) and post training (~instructions and RL)

Reflection isn't really "from no where", they have huge financial backing, 10K of the latest GPUs, and many ex leads from the established labs.

There are many measurable metrics and I don't see any that seem impressive. Care to share the ones you found exciting?

Thanks for what you've done and continue to do for the ecosystem!

Unfortunately Nathan recently left AI2, along with some of the leadership most involved in their truly open models. https://www.geekwire.com/2026/allen-institute-for-ai-ceo-ali...

Yes, but even nvidia doesn't have unlimited gpus. Everything they hold for internal teams is a reduction in revenue and customers can have aggressive commitments.

This is one of the jagged mismatches between users and LLM developers. The median user doesn't care or want to know about static models and knowledge cutoffs and whether a model can do tool calls or if tool calls even happened. They just want something that works.

Fortunately increased capabilities seem to make this a basic expectation with new releases.

It can be quite hard to determine what needs a tool call or not. LLMs are not well calibrated to what they know and don't know, and tool calls can add latency and extra costs. There are lots of things that are "obvious" right until they aren't - especially political events and disasters.

Clem knows they have the community support and potential legal leverage here. It's not unreasonable to ask for a lot. It's in everyone's interests to play nice. Huggingface could do a lot with more compute, even with nvidia backstop. Anything with large scale gpus quickly gets into ~year lead times, "contact sales", and complex talks. Large scale being only > 64!

If you take the recent anthropic misuse report seriously, they don't need to "take over weapons" because they're actively being used to give a lift in capability to groups who already want to use weapons.

> The six cases in this section are divided into two parts. Part I covers four cases in which the actor in question used Claude to develop software for weapons themselves: a guided rocket program, in which the actors conducted a live field test; a design and proposal work on a system to intercept torpedoes; software for a drone swarm, tested in simulation, with its code loaded onto real boards; a targeting software for electronic warfare and for suppressing air defenses.

https://www.anthropic.com/threat-intelligence-report-septemb...

This is likely being discussed because it's relevant to this similar modern day story involving Tesla: https://news.ycombinator.com/item?id=49686766

Sure. But that was months ago, and not the newsworthy event scoped to "The Last 24 Hours" as the title says :)

> Maybe you’ve already heard about the guy who resigned from OpenAI.

While he worked at both OpenAI and Anthropic, he resigned from Anthropic. Mistaken reporting in the first few sentences, definitely a horror concept.

That might be true, but nothing else has been as effective at accelerating model development and research sharing.

In earlier circles they were known as the "pytorch-pretrained-bert" guys, still under the huggingface company name. IIRC it was a health chatbot type startup.

"Improve the model for everyone" can be implemented in so many ambiguous ways.

https://news.ycombinator.com/item?id=49643513

Reminder that there are degrees of "trained on conversations". From John Schulman:

> pretrain on user data, with users' tokens as prediction targets: high regurgitation risk, improper

> use user prompts to distill large models into small ones: low regurg. risk, some companies probably do this

> use user traces to construct RL tasks: low regurg. risk, because RL has low memorization abilities, but can extract customer IP, depending on how it's done. Ranges from benign "use explicit user feedback in reward model training" to invasive "upload user's coding environment and commit history to turn into rl envs"

source: https://x.com/johnschulman2/status/2097440545853637108

The canary string was more about inadvertent scraping or analysis in other papers. Not direct training on user data. And the use of BB has eroded quite a bit, with BB-Hard or other variants being typically used.

Commenters in these discussions always confuse wikipedia editors with Wikimedia employees - this is the latter!

Folks are concerned that nvidia won't support these efforts if it gets models running on competing hardware. Two responses:

- The projects started without HF/nvidia involvement and were massively successful BECAUSE folks want to run models on their own devices.

- With highly capable agents, "hard to implement" should be less of a barrier. The inference market should in some sense become more efficient, as agents should make it easier to transition between software and hardware solutions. Sure there might be less training data for integration platforms, but if we've learned anything in the past few weeks, it's that agents can be remarkably persistent.

Care to comment on this model from a quite serious company being labeled as derived from GLM?

Also, do you have a source for: "99% certainly european chips, almost certainly EIC-funded. (Eg. Hailo, Axelera, ..)"

For Hailo, the best I can find is https://hailo.ai/products/ai-accelerators/hailo-10h-m-2-ai-a...

Which states that this is a 40 TOPS int4 PCIe device, not what you'd put in a datacenter and certainly not what can run a frontier class model (or the model above). Embedded devices for inference are really cool! But that's the opposite end of scale for what goes into a data center.

On Artificial Analysis it's listed as "Quasar 438B (max, based on GLM-5.2)" - so you see exactly right. Not sure if this was changed post publicity drive or not, this is the first I'm seeing about this model.

https://artificialanalysis.ai/models/quasar-438b