Sabitlenmiş Tweet
Stijn
3.9K posts
Stijn
@StijnSmits
Fouding AI Engineer @ Schematik ex-Zyphra 2x marathoner
NO/NL 🇳🇴🇳🇱 Katılım Mart 2011
1.4K Takip Edilen1.9K Takipçiler
@cheatyyyy That would make sense on the serving-cost side.
How are you estimating the <50B active params part? Is that mostly from serving-cost/latency intuition, or something more specific?
English
My hunch: if 5.5 Instant shares the same base model as 5.5 Thinking and is being rolled out broadly, then one or more things must be true:
- it’s not an insanely large model (1-3T?)
- the Instant path is heavily optimized/routed/constrained
- OpenAI has a lot more inference compute available than people assume
OpenAI@OpenAI
GPT-5.5 Instant is starting to roll out in ChatGPT. It’s a big upgrade, giving you smarter, clearer, and more personalized answers in a warmer, more natural tone. And it's also more concise, which we heard you wanted. We think you'll love chatting with it.
English
🚨🚨🚨- Internet Friends v24
Alright alright lads & lasses, we're on again!
When, what, where, who 👇
📅 7th of May - Thursday!!
🕐 18:00
📌 Internet Friends HQ - Jacob van Lennepstraat 78H
🧑💻 30-40 ultra ultra nerds talking VScode and all that
Format: No format + 🍕 + 🍺
Ping me if you're joining 🫡
English
has a 12M token context window
benches at frontier level for coding
uses 1000x less compute
if this is real then it is a genuine breakthrough and this company is about to make a lot of money very fast
Alexander Whedon@alex_whedon
Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens - Less than 5% the cost of Opus Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention). Only a small fraction actually matter. @subquadratic finds and focuses only on the ones that do. That's nearly 1,000x less compute and a new way for LLMs to scale.
English
@MaximeRivest Classic network effects. More extensions attract more devs who build more extensions. But I wonder - at what point does malleability become fragmentation?
English
Pi is incredibly malleable, to point that I am starting to believe it is something to study to understanding software extendability.
Does anybody know: why is pi oh so malleable? why is pi so easy to make extension for and to adapt to ones own needs?
English
My final Zyphra project: dimensional reduction on the parallelism mesh
Zyphra@ZyphraAI
Introducing folded Tensor and Sequence Parallelism (TSP), a new way to split large models across GPUs that achieves lower per-GPU peak memory than any standard parallelism scheme. Scaled on @AMD MI300x. Bigger models, longer contexts, and higher throughput 🧵
English
big news: @StijnSmits just joined @schematikio as founding AI engineer :) he has 400k+ downloads on hugging face and now he's helping us building the sickest hardware ai agent in the world, woohoo!
English
English
English
English
@SamuelBeek @StijnSmits @schematikio Congrats to both of you!
Can't wait to just prompt my PCBs into existence :)
English
English
