Ed25518

caffeinate claude

[Ed25518 @Ed25518]

@BlackRock blackrock digital assets is packed with degenerates who don't even care to double-check "research they put out" and are actively robbing pensioners

[Ed25518 @Ed25518]

@0xSero nice try fed

[Ed25518 @Ed25518]

@benhylak god that glass pannel is gorgeous

[Ed25518 @Ed25518]

@scheemunai conveniently left out o1, o3

[Ed25518 @Ed25518]

@andrewmccalip YOU DONT SEEM TO UNDERSTAND? THIS IS NOT A PURE LLM! IT IS A NEUROSYMBOLIC HYBRID ARCHITECTURE. PREDICTING THE NEXT TOKEN CANNOT GET A NEURAL NETWORK TO DO CAD. AUTOREGRESSIVE TRANSFORMERS CANNOT ACHIEVE THIS, IT NEEDS A NON-GENERATIVE WORLD MODEL (grep and context compaction)

[Ed25518 @Ed25518]

@kvablack >135.1GB GPU memory

[Ed25518 @Ed25518]

@sneaknsneak i hate the third world as much as the next guy, but processed forms and extracts are illegal in most countries.

[Ed25518 @Ed25518]

@GeoffreyHuntley @thsottiaux FYI, if you have the money, you can still use codex like you would with a subsidized plan by getting an API key

[frankie steins gay uncle @ewkewpid]

i love you leopardus tilcayo

[Ed25518 @Ed25518]

@DanishMFA how come you use legal style (very poorly, but anyway) when they arrest protesters thousands of miles away in a country you barely have any connection with, but you send a black and white sans-serif plaintext when a superpower threatens to take away your land and citizens

[Ed25518 @Ed25518]

@bycloudai hate to break it to you but it's a transformer. probably decoder-only too.

[Ed25518 @Ed25518]

@NielsRogge considering that JSON classification was done by an autoregressive language model for the most part since 2023, this is pretty good news.

[Ed25518 @Ed25518]

@sanlsrni its not a diffusionmodel and its definitely not diffusionpilled

[Ed25518 @Ed25518]

@max_paperclips @Cyndesama https://t.co/lj1YKDlod5
https://t.co/ACjXBWwv0F
https://t.co/Q2eOs603cz
https://t.co/wKBZ8TYkBQ
https://t.co/9DJ0F3tjWT

[Ed25518 @Ed25518]

@max_paperclips https://t.co/5YUiaS9XxH

[Ed25518 @Ed25518]

Is it a smaller LLM? No, but actually, yes. It's not an LLM, but it's almost certainly grown from one. The reward model in RLHF is a transformer with the LM head chopped off and a scalar head bolted on. Their system is probably this but generalized. Many heads, typed answer spaces, calibrated distributions instead of a raw scalar, and trained as the product rather than as a training-time side tool.

- Pareto-frontier intelligence almost certainly comes from distillation.

- Output tokens are free because there are no output tokens. The "output" is a softmax per question, essentially zero compute.

- Input at $0.042/MTok is small-model prefill pricing. Prefill is compute-bound, batched matrix-matrix work at near-peak GPU utilization. Autoregressive decoding is memory-bandwidth-bound and sequential. If you only ever prefill, you're using the hardware the way it wants to be used, which is what "hardware-aware parallel sampler" means.

- 70–500ms latency is one forward pass over a few hundred to few thousand tokens plus the network hop. Nothing else fits that window.

- Cardinality capped at 255 is a uint8. The answer index is literally a byte. Above that they score candidates independently and then pick, which is what you'd do if the head can only address 255 options at once.

- "Mathematically can't make type errors" just means the output layer is defined over the schema. There's no string to parse because there's no string.

[Ed25518 @Ed25518]

Is it a smaller LLM? No, but actually, yes. It's not an LLM, but it's almost certainly grown from one. The reward model in RLHF is a transformer with the LM head chopped off and a scalar head bolted on. Their system is probably this but generalized. Many heads, typed answer spaces, calibrated distributions instead of a raw scalar, and trained as the product rather than as a training-time side tool.

- Pareto-frontier intelligence almost certainly comes from distillation.

- Output tokens are free because there are no output tokens. The "output" is a softmax per question, essentially zero compute.

- Input at $0.042/MTok is small-model prefill pricing. Prefill is compute-bound, batched matrix-matrix work at near-peak GPU utilization. Autoregressive decoding is memory-bandwidth-bound and sequential. If you only ever prefill, you're using the hardware the way it wants to be used, which is what "hardware-aware parallel sampler" means.

- 70–500ms latency is one forward pass over a few hundred to few thousand tokens plus the network hop. Nothing else fits that window.

- Cardinality capped at 255 is a uint8. The answer index is literally a byte. Above that they score candidates independently and then pick, which is what you'd do if the head can only address 255 options at once.

- "Mathematically can't make type errors" just means the output layer is defined over the schema. There's no string to parse because there's no string.

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Ed25518 @Ed25518]

the problem with the Higgsfield After Effects harness / skill is that the greatest issue in automating motion design is After Effects itself, and most problems go away if you just decide not to use it at all.

[Higgsfield AI 🧩 @higgsfield_ai]

what if it actually can?

[Ed25518 @Ed25518]

@pansexualDBZfan this only works if you live in a gated community

[Ed25518 @Ed25518]

@TerribleMaps holy based czechslovakia

[Ed25518 @Ed25518]

@luckenco @Laz4rz you could also try https://t.co/WkDaPgFvln perhaps