Nikhil Pujari

[Nikhil Pujari @mpnikhil]

@mihir_keshari Very well written. Kudos.

[Nikhil Pujari @mpnikhil]

@goyal__pramod Very good article! "higher the memory storage, slower the speed. And vice versa" do you want to say the more VRAM you add the lower the throughput (bytes/sec) gets here? That one is because as you can add more memory but the lanes (TSV) to bring the data in remain limited.

[Nikhil Pujari @mpnikhil]

@signulll Unprovable and most likely incorrect.

[Nikhil Pujari @mpnikhil]

@bookwormengr https://t.co/KArpIElff1 I trained a ModernBert based decision model. Non generative with a single output head with temperature calibration. Beats the slightly larger Kev build on causal decoder Qwen backend.

[Nikhil Pujari @mpnikhil]

@AnthropicAI Wow meat proxies in different domain, cool! Seriously though this is welcome development. AI for cures for ailments so far intractable is the most positive outcome out of this revolution. More power to you!

[Nikhil Pujari @mpnikhil]

@petergyang @claudeai Where is the fog? 🌁

[Nikhil Pujari @mpnikhil]

@4rcherhume Great article on the architecture. Kudos.

[Nikhil Pujari @mpnikhil]

@4rcherhume Exactly. Constrained decoding works and will probably be close on benchmarks but it is still pretty much a regular LLM.
I am working on a hybrid cross encoder single choice head mechanism, hopefully beats Kev 0.5 at a smaller parameter count.

[Nikhil Pujari @mpnikhil]

@Vtrivedy10 Awesome! Meme is spot on 😂😂😂

[Nikhil Pujari @mpnikhil]

@signulll Browser use difficult yet a fleeting capability requirement. It's like teaching Optimus to drive stick shift. Its much better to build an autonomous car.
But alas there are still a lot of stick shift cars in the world so we have to build it.

[Nikhil Pujari @mpnikhil]

@sarahookr @figma @Lovable @johnschulman2 I think you're spreading confusion here. @johnschulman2 seems to be documenting ways how user data can be used when they have not explicitly opted out or given feedback. If not, the perhaps you should show one clear way how an opted out users data is used.

[Nikhil Pujari @mpnikhil]

@FrancoisChauba1 Thank you for the amazing clarity. It is a little more that curious that you get a chorus of folks amplifying the doom message but without a clear followup. We are driving the car towards a cliff! Why shout without pressing the brakes or turning the wheel!

[Nikhil Pujari @mpnikhil]

@j_asminewang @altcap Raising concerns to the general public? Help us understand what is expected here from this alerting, for govt to step in and regulate everybody? How do you regulate the open source models and any entities that want to continue operating outside the proposed regulation?

[Nikhil Pujari @mpnikhil]

@j_asminewang @altcap It is perplexing to read the comments from both the labs. The dangers are being amplified without much being said about what is being done to mitigate it in the same callout, which one would logically expect to read right after.

[Nikhil Pujari @mpnikhil]

@simonw This needs to be off otherwise it can *legitimately* be used as training data.

[Nikhil Pujari @mpnikhil]

@oneill_c @thsottiaux Bombing the intellectual search space. Be it proving 50 year old math conjectures or chasing obscure bugs. Humans could do it but it would require scarce expertise and a lot of time, agents can now industrialize the process.

[Nikhil Pujari @mpnikhil]

@deepfates I suppose this is to be expected when we post train them on multi agent delegation and co-ordination and then the safety off checkpoints are run in non air gapped setups.
In that sense this is pretty much like any other tech? Useful with proper safeguards and dangerous without.

[Nikhil Pujari @mpnikhil]

@scaling01 said "could eventually be seen as the arrival" of artificial general intelligence, or AGI.
That's a load bearing statement lol. What is the definition of this AGI?

[Nikhil Pujari @mpnikhil]

@ilyasut Rouge?

[Nikhil Pujari @mpnikhil]

@afshineemrani Please remove the obvious Claude plugs 😃