when I went to korea I was in shock at how fast the payments were, it's like it auth'd before you finish tapping... same vibes w/ gemini 3.8 serving right now, done before you blink...
maximum in group signaling if we could design a shirt around this concept:
spark.sql.shuffle.partitions 200
random forest from first principles
one of my fav talks at Neurips this year was about mode collapse in LLMs and its societal impacts. This has recently motivated many arguments on this platform on how LLMs can never discover anything new so don't outsource our creativity to them..
@liweijianglw
another example of mode collapse, but this time not from an LLM
teapawt.org/jobergum/status/210297…
when one intent dominates the distribution (a toxic platform), earnest inputs get mapped onto that intent: many possible readings -> one familiar interpretation -> a defensive response
stealing this for inteview question:
You have an encoder-only language model like BERT. You want to classify text into categories that weren't present during training. You can't fine-tune the model when new categories appear. How would you implement zero-shot classification?
NLI head on top of BERT??? why not? very a pro pos given DS brought back T5 this week too...
Jev looks cool. I hope to try it.
But so far, the most interesting thing to me about it is the reactions. They suggest a lot of folks are not aware that encoder-only classifiers exist, have existed for a long time, and are good and valuable and widely used.
It makes sense. LLMs have become so powerful and widespread so fast, it's all that some people have seen.
For instance, for years now, you've been able to fine-tune a ModernBERT model (or maybe LFM2.5?) to do fast cheap structured output from a language input. I expect Jev is much better, and IIUC does not need training, but the category is not new.
Jev might be what you'd design if you took DSPy's Predict(Signature) abstraction seriously enough to build a foundation model specifically for executing it.
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
the resurrection of T5/BART was not in my 2026 bingo card
🧠 Asymmetric architecture. More intelligence, less cost.
🔹 552B-parameter MoE.
🔹 New Causal Encoder–Decoder architecture: just 8B active parameters for input, 16B for output.
🔹 New pre-training methods + larger-scale RL post-training deliver benchmark results ahead of flagship models, including DeepSeek-V4-Pro.
2/6
