Imran Qureshi

sharing useful science and AI ideas

[Imran Qureshi @explorations_iq]

@LangChain Jev could be very useful for scalable oversight where we have strong models produce evidence for their claims and Jev like models could be the weak judge

[Imran Qureshi @explorations_iq]

@Steve8708 But all of these demos are real. Sure its not generating stuff "from scratch" but even LLMs choose tokens from a predefined set so this video is actually misleading

[Imran Qureshi @explorations_iq]

@DXhusni Nice, what software did you use for the virtual robot arm?

[Imran Qureshi @explorations_iq]

Trying out Jev's drawing capabilities: its very impressive it was able to do some spatial reasoning with my early / naive attempts. Also it only cost ~$0.0029 (video sped up 2x)

[Diogo Almeida @CompleteSkeptic]

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

[Imran Qureshi @explorations_iq]

@jackcheng This is an amazing use case, I can imagine an HCI across any application enabling speedy voice-app interaction. You just need to set up the action space properly

[Imran Qureshi @explorations_iq]

@nutlope Super cool, but I wonder how does this compare with embeddings or NLP techniques. I'll have to do some benchmarking

[Imran Qureshi @explorations_iq]

@vinnylarouge Very nice, have you tried any evals with similarly sized models against jevlike?

[Imran Qureshi @explorations_iq]

@CompleteSkeptic Very cool! Is this a discrete/masked diffusion model under the hood (constrained to a fixed schema per slot), or something else entirely?

[Imran Qureshi @explorations_iq]

@alexcooldev Even so, how do you know one of your shots is going to be reasonably good at succeeding?

[Imran Qureshi @explorations_iq]

@shiri_shh Princeton has a intro to robotics course that is really great so far: https://t.co/i08jkuj9hz

[Imran Qureshi @explorations_iq]

@ZentrixHQ Very cool use case but how do you know it's all accurate?

[Imran Qureshi @explorations_iq]

@iChuloo This is a great idea. I am making one that uses Minimax H3 to have video calls with native speakers that you could talk with live for various scenarios. How is the cost like per conversation?

[Imran Qureshi @explorations_iq]

I've been building a tennis practice rig that synchronizes a ball machine + phone camera + computer vision all running locally on device. It can track my position, swings, and ball and feed balls creatively based on my position.

[Imran Qureshi @explorations_iq]

@skalskip92 You could also use Astra for labeling to create small-purpose models that are much more efficient and faster than Astra is

[Imran Qureshi @explorations_iq]

@DilumSanjaya Very cool, this would be a great teaching tool if you could remove a part and see what happens or have the user build it themselves piece by piece.

[Imran Qureshi @explorations_iq]

@gregisenberg I'm assuming the liability issues would become a problem, because if a part breaks you're still on the hook, which might be the hardest part of this business. Otherwise, if you know the business, some amazing designs are possible

[Imran Qureshi @explorations_iq]

@umesh_ai Another nice addition would be the shadow of the boy grown up interacting with the tree's shadow. Nice idea though

[Imran Qureshi @explorations_iq]

@suemaruuuuuuX Now I wonder if you could take a art piece like the Mona Lisa and have it draw it layer by layer? Although I guess it can only operate in specific pixels, not strokes, so perhaps something for the next GPT.

[Imran Qureshi @explorations_iq]

@FrancoisChauba1 This should be a standard tool for all LLMs to be able to call a stronger LLM for planning and strategic purposes. I think it's a great idea!

[Imran Qureshi @explorations_iq]

@ashebytes Imagine doing this with every patent in existence