Seva

i write cool stuff
tech lead | web dev | ai/ml

[Seva @Two_Above]

@kalomaze This is incredible dude

[Seva @Two_Above]

@highyieldYT I’ll never financially recover from this

[Seva @Two_Above]

After burning through a few hundred million tokens with Opus 4.8 I can say for sure that it’s a lot better than 4.7

[Seva @Two_Above]

It will be very interesting to see how it does on DeepSWE

[MiniMax (official) @MiniMax_AI]

Introducing MiniMax M3: The First Open-Weights Model to Combine Three Frontier Capabilities

- Coding & Agentic Frontier: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas
- MiniMax Sparse Attention scales context to 1M
- Natively Multimodal from Step Zero

API: https://t.co/fHRdSV7BwZ
Token Plan: https://t.co/BDCycxepZw
🚀New! MiniMax Code: https://t.co/GvB4YiB6Ul

Weights & Tech Report in ~10 Days

[Seva @Two_Above]

Does codex goal mode have a max run time? I hit 24 hours and it stopped :(

[Seva @Two_Above]

This is scary impressive coming from GPT 5.5. Crunched though very convoluted decompilation and RE task

[Seva @Two_Above]

Getting good vibes from this model, both on work and personal projects!

[Claude @claudeai]

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.

Available today at the same price.

[Seva @Two_Above]

/model claude-opus-4-8 if you want to use it early

[Seva @Two_Above]

How I sleep knowing that world records are records for a reason

[Seva @Two_Above]

@WallisDev Can confirm. I’ve had literally ZERO issues on us-west-2

[Seva @Two_Above]

@berrytop9 Diet Coke is better for this btw

[Seva @Two_Above]

@xyster Intel is goated until they stop playing catch up. Can’t wait for their C series cards

[Seva @Two_Above]

@JustJake Yeah it was a dream of mine to get my hands on one of these. Intel would have made a killing post~2025 if they stuck with them for a little while longer.
Perfect disk for dev work and dbs

[Seva @Two_Above]

@lauriewired So this is in a way more usable than Intel optane set up as swap?

[Seva @Two_Above]

Look how much compute is being left on the table. We can 10x LLM speeds if needed

[zR @zRdianjiao]

🚀 GLM-5.1-HighSpeed is live: 400 tokens/s — a new speed ceiling for flagship-tier LLM APIs.

Not a smaller model traded for speed. A flagship from @Zai_org that's also the fastest.

📖 Full technical deep-dive 👇
https://t.co/nLEFdMf2Ea

[Seva @Two_Above]

checked on onedollarstats and holy - noitool is doing really well. so proud of it

[Seva @Two_Above]

@xyster Thanks!
And totally valid use for heatsinks

[Seva @Two_Above]

@JustJake Sane solution to an insane incident
Thanks for the transparency
Can you share any info what the reason for the suspension was?

[Seva @Two_Above]

Please go read this - really great paper with some fancy and clever training and architecture decisions

Here’s some cool things in no particular order:

HRM-Text omits broad raw-text pretraining and trains exclusively on instruction-response pairs from scratch.

MagicNorm, which ex-ploits the asymmetry between the forward and backward computational horizons induced by truncated backpropagation through time.

combining small recurrent reasoning models with external or learned knowledge stores is a promising direction

We hypothesize that the instabilities observed under deep BPTT in looped architectures are a consequence of the intrinsically multiplicative structure of gradient propagation through repeated iterations. Specifically, gradients backpropagate through products of Jacobian-like operators across loop steps, and theory for products of many random matrices predicts that the logarithm of the norm of such products is approximately Gaussian, implying lognormal-like variability in gradient magnitudes and increasing separation between typical and extreme values as backward depth grows

[Guan Wang @makingAGI]

The HRM-Text paper is now available 🎉

HRM-Text explores a different approach to language model pretraining: hierarchical recurrent computation, task-completion training, and latent-space reasoning.

At just 1B parameters, HRM-Text achieves competitive performance with dramatically lower training cost and data requirements.

1B parameters
40B unique tokens
~1 day of pretraining
~$1000 training cost