Sahebjot Singh

AI & Search

[Sahebjot Singh @zereraz]

@Lon @simonw So it reads the config.json, adapter_config, safetensor headers of huggingface models and turns it into an interactive TUI viewer.
I was training some LoRA adapters and wanted to see where they were fitting.

Was a temporary tool for work but I'll open source it :)

[Sahebjot Singh @zereraz]

not llm as a judge, rather llm in a harness as a judge

[Sahebjot Singh @zereraz]

Thanks to GLM 5.2 we know MoE is the right direction

After Qwen 3.6 27B it was getting confusing...

[Sahebjot Singh @zereraz]

@LottoLabs shouldn't it be Dell Pro Max with GB300 ?

[Sahebjot Singh @zereraz]

When china commoditizes coding models, the next big value add has to be in Science.

[Claude @claudeai]

Introducing Claude Science, a new app designed with every stage of research in mind.

Artifacts traced to their code, environments managed on demand, and 60+ optional scientific databases that you can connect.

Available now in beta.

[Sahebjot Singh @zereraz]

so we already got opus tier trained on chinese chips...

[Sahebjot Singh @zereraz]

@difficultyang kyber rack

[Sahebjot Singh @zereraz]

and at the same time ability to further design and manufacture power tools for knowledge work

[Sahebjot Singh @zereraz]

llms are power tools of the knowledge work era

[Sahebjot Singh @zereraz]

Gov. telling what models to sell / limit, is just step 1...

Sort of a new dynamic, state-AI partnership an unprecedented alliance.

[Sahebjot Singh @zereraz]

@Dorialexander Someday NVFP4?

[Sahebjot Singh @zereraz]

@natolambert kimi 2.5 was a moment as well, multi modal and fast.

[Sahebjot Singh @zereraz]

Models can absorb new skills given we have enough good envs to do RL.

Who knows which skill will complement which one, composition of skills emerge.

Ofc some signal has to exist between the model and the envs we are training it with.

signal - non zero reward variance

[Sahebjot Singh @zereraz]

@blader benchmark results + weights release

[Sahebjot Singh @zereraz]

maximise cumulative reward <> minimise kl divergence

[JFPuget 🇫🇷🇺🇦🇨🇦🇬🇱 @JFPuget]

I've warned about this many times. The brain is like a muscle. If you don't exercise it, then its strength decreases.

[Sahebjot Singh @zereraz]

yet every query into the GPT model is an Exploration

[Sahebjot Singh @zereraz]

First search had to exist as Google on internet for compression to take place into GPT models

Exploration to Exploitation

[Sahebjot Singh @zereraz]

@TheAhmadOsman let's not forget deepseek r1