foobar10000

Fixing healthcare or economic policies is typically an act of congress type of thing - not even an amendment. Large popular support - and it can be fixed. Fixing freedom of speech or the ability to vote at all - once it is gone - usually takes a price payed in blood. Sometimes rivers of it.

I hate to use Claudish here, but "postnatal" and "dead" are quite load-bearing here :) Clinically dead? Coma? Born human baby, but with anencephaly - expected lifespan is days, and there's literally no brain - but, well, are they a person? Ask 1000 people - and I'm betting we'll get a nice bimodal distribution.

I take your point that a human walking around a city will - out of that 1000 responses - hopefully get 990+ of "yes, a person" - but that will start to degrade quickly once you hit boundaries.

And the less said about dehumanizaion - untermensh and that entire line of thinking - the better.

While I was not participating actively in the discussion - I just want to thank both participants for embodying the spirit of HN and academic discourse - especially in the face of a difficult debate that stirs strong emotions.

@svnt, @hnlmorg - this conversation - and the links it produced - are a fascinating read, and honestly gave me hope for discourse on the open internet.

Unfortunately not satire - modern models can be trained to be better than humans in recognizing emotions straight off camera.

Sigh... I'm willing to bet (and this is not a hard thing to bet on) that even bb(50) is independent of ZFC and similar, and Omega(1KB) is > 748 , which is at this point known to be independent of ZFC.... Please do not ask the LLM to break the universe :)

Citadel already is. They are happy to do it - bedrock is the host of choice. They are less happy about cost - but that is a question of value :)

One - and I do not mean to be snarky - you can literally ask Gpt 5.6 Sol this - and if you want to see cool stuff - Fable running in their app (not website) has a view thinking button that is actually a good way to explore the adjacent fields, etc.

The non-sofic group one is definitely a big deal - would have been a Fields medal if discovered by a human.

Glm 5.2 nvfp4 on 4 b300 with dpattn 4 and ram will get you about 20 users live at 400k context - and 60 easily if you give that server 2tb of ram and 4 nvme 8 tb drives. There are some sglang patches needed - but we will be releasing them soon.

Or more colloquially : paperclip maximization . From OpenAI - you know, the guys who _really_ know this... Sigh... Did they finish the prompt with "And do whatever you can to get this done!" ? Cause that's the only thing that would make this even dumber...

This is an _amazing_ typo :) Thank you, thank you :)

3T at nxfp4 (which is most of it) is only 1.5TB of vram - so 8x288GB B300 or MI355 will do it if you are careful with context - maybe dp-attn? Certainly not TP. 2 of those together can easily serve it. The new AMD MI400 are at 400GB+ each, so 8x of them will nicely fit with KV to spare.

Don't forget that you are not really seeing the thinking tokens used - so non-trivial to count them.

Yeah, if you have a fixed llm topology, you can just effectively burns 2 top layers of the chip as Rom (model weights) - which has a per area density even better than dram - so it’s just attention and kv streaming that is hbm to sram transfer.

Most big model weights will not fit a single reticle sized chip - so you’d have prob 30 different chips to split the model .

And you’d need super fast chip to chip comms for the all-reduce and similar.

So scaling to 1T models is hard - and a long lead time - but can be very power efficient.

Well, for a lot of agentic stuff nowadays, having 250k-500K context is where things live - and the benchmarks don't really show that unfortunately - but they could :)

Agreed - there was always a set of things I wanted to do that I knew the magic core for, but wanted a team of implementers for the curft, the 100k of actual testing harnesses, hyperparameter exploration, etc.. . I now have that team of implementers. All the problems seem research-y though - optimal binary transport systems that are zero-copy and compatible with languages, fast physical simulation optimizers, etc etc... So, things that all had a _LOT_ of busywork around the magic core.