Paul Bridger

ML Consultant and world traveller 👽
Former ML Optimization Lead and Advisor @WOMBO🌈
Former Technical Cofounder @Gluent 🤖 (sold to Google)

[Paul Bridger @paul_bridger]

New article! Object detection at 2350 FPS using 8-Bit Quantization and @NVIDIAAI TensorRT, from a Pytorch model. https://t.co/kcN3iUvmqA

[Paul Bridger @paul_bridger]

@elvissun Canonical top-shelf texts for a given subject.

[Paul Bridger @paul_bridger]

@elvissun Yeah, I did this and it was amazing. Orchestrated by CC, using Gemini TTS, and driven by top-quality sources from a notebooklm.

[Paul Bridger @paul_bridger]

@Prince_Canuma That’s the level up screen

[Paul Bridger @paul_bridger]

@vtchakarova Chart says 12%, not 22%, in the conservative scenario.

[Paul Bridger @paul_bridger]

@levelsio Give it more specific context? Clearly if you ask without context you get a normie answer. With a unique context you get a normative answer for that context, so should be more unique.

[Paul Bridger @paul_bridger]

@nearcyan motorcycle + music
so thrilling people kill themselves getting too excited

[Paul Bridger @paul_bridger]

@nathanbenaich Dojo

[Paul Bridger @paul_bridger]

@fchollet @ylecun Optimus Prime would disagree

[Paul Bridger @paul_bridger]

@gregeganSF I like how you turned this maths problem into a little story. You should consider writing books!

[Paul Bridger @paul_bridger]

All my favourite songs are called “Untitled (Remix)”
@SunoMusic

[Paul Bridger @paul_bridger]

@gregeganSF Classic ChatGPT glazing, hallucination and lies. Trusting ChatGPT is user error at this point. Try a serious LLM like Gemini.

[Tim Dettmers @Tim_Dettmers]

Just to clarify this benchmark. This is an apple to oranges comparison.
- Cerebras is fast for batch size 1 but slow for batch size n.
- GPUs are slow for batch size 1 but fast for batch size n.

I get >800 tok/s on 8x H100 for a 405B model for batch size=n.

Cerebras' system is great for batch size=1 inference, but usually, this is only effective for personal use or latency-critical applications since it has too high a token/$ cost for deployments.

Some people point to agents, but agents usually also need to do parallel calls over multiple data points. With my open-source-derived system for agentic tasks, I can get >100 ktok/s across 128 GPUs for Qwen 32B.

I want to point this out because people in the comments seem confused or skeptical. The numbers are real, but the comparison is not transparent about the batch size differences.

This is common in many other marketing benchmarks, so when you see big gaps like this it is usually not a fair comparison.

[Paul Bridger @paul_bridger]

From SQL to AI: Modern LLM prompting is declarative programming reborn. 'Give me a CSV with these columns...' is the new SELECT statement. Define your output structure, let the model work out the implementation.

[Paul Bridger @paul_bridger]

Thanks Parshant, I just had to explain lowkey goated to my wife

[ΩxDrP @parshantdeep]

1. Keep answers short, sweet, and concise.
2. Provide insightful and "lowkey goated" responses.
3. Make answers first principled:
- Break concepts down to fundamentals.
- Build up from these fundamentals for every query.
4. Follow all previous instructions while implementing the first-principles approach.

[Paul Bridger @paul_bridger]

https://t.co/ElmitvPdmd for @Neovim is awesome.
I hated the UI and workflow in Cursor - reaching for a mouse feels like failure.
Thanks @yetone and @aarnphm!

[Tanel Poder 🇺🇦 @TanelPoder]

Talked to my friend @paul_bridger who works in ML/AI, he was spot on, saying this:

“AI will not replace people, people with AI will replace people without.”

[OMEGA @omegalabsai]

OMEGA Labs is holding our first ‘Garage AGI’ series on Spaces where we dive into the humanity of AGI development. We’re betting big on Bittensor and @parshantdeep is going to discuss why with @_sshahid_ and @paul_bridger

https://t.co/fRw2n089Wq

[Paul Bridger @paul_bridger]

@giffmana Even better would be “Run all in parallel”

[Paul Bridger @paul_bridger]

@JohnDCook Just experience the moment, damnit.

[zion | w.ai @benzion_b]

According to this, @WOMBO Dream would account for >10% of all generated images 👽