@ditpoo
Account based inIndia
About this account
- Account based in
- India
- Connected via
- India App Store
Account-level information from X, not a live location or the device used for a specific post.
Curious, OpTi-MaX CS, AI R&D and Engineer. Simulators & Emulators AI Email: [email protected] , [email protected] Github: https://nitter.cf/t.co/LjEONGCZBw
India
Joined January 2013
- Tweets5.5K
- Following2.1K
- Followers96
- Likes36.7K
My Core contribution temporal (optimax scaling) optimization, unlocking full computation (generation) capability of AI model and along with that proving that we might be living in a simulation.
The temporal optimization of 2x slow motion increasing generation (computation) of video with respect video generation model, along with other optimizations, proves digitally about temporal nature of simulation and of reality.
Also same should apply physically in physical reality, i.e real world, eg. there might be temporal component with respect to fluid dynamics, simulations, eg. navier stokes
nitter.cf/ditpoo/status/19824242…
Temporal Optimal Video Generation Using Grandma Optimality.
Universal optimax prompt enhancement for computation eg. image, video generation. If the model can be tuned or prompted to work in this way.
"Make the video 2x times slower speed, slow-motion, while maintaining all the visual elements, quality and aesthetics."
Test for your self. Samples Below, this is the secret sauce behind getting super high quality generations out of same models as compared to simple prompts or simple way to do it.
Like the forward thinking of investing in and building the naval fleet, that helped city of athens survive, when athenians were at crossroads with their wealth and technology.
AI too is a similar forward looking bet and a crossroad that humanity has to make the right choice over, to ensure a brighter future, and a greater chance at surviving what ever that universe throws at humanity,
For greater economic prosperity via expansion of economy, for greater discoveries and inventions, for solving all the toughest problems faced by humanity, we just have to make the right call with this one, AI.
If by vision of a single leader in “Themistocles bets everything on the navy”, can save athens, so can we (those who can feel and see the future in AI) by taking bets and going all in it.
- Jev is as an example of “System One” like AI: fast, cheap, lower-intelligence outputs instead of slow, high-reasoning ones.
- That pattern (system one not jev) could matter for agents, tool calling, computer use, and physical AI, because those jobs need many cheap, fast decisions.
- Jev’s interesting part is speed, cost, and code integration — not replacing current models on capability.
- It's only usable in space with relaxed correctness, reliability requirements and that of lower intelligence or cognitive abilities which makes it's use quite limited.
- But it's a exciting glimpse at future of system one mode AI.
On Jev & System One Thinking Models.
I don't know if it's (jev) a single shot structured generation as parallel gen (maybe diffusion llm) via a smaller sized model that is leading to such numbers (speed and cost), but it's a good way to explain system one thinking modes of AI models we might need or see in future.
What we see with examples or use cases of jev is a good example to explain system one thinking modes of llm's, these are super quick, cheap and lower intelligence outputs or actions done by AI.
These have great implications for agentic tool calling, orchestration and computer use, as those will become this fast and this cheap in future, same applies for physical AI (robotic/physical actions and inference) like FSD's (full self driving systems) via system one thinking mode of AI models (llm's) that can work in parallel to normal thinking or other processing.
Great example of this are interaction based duplex models that can do both voice interactions and thinking in parallel, simultaneously super fast (in real / speech time) i.e real time speech and thinking, now imagine that like jev with agentic tool use, computer (also system and env) use and physical AI (physical actions and use cases).
With decision making of higher cognitive and intelligence requirement, current AI tech can't do even frontier level intelligent decision making at system one thinking level, but in future that might be possible, so that is wrt to decision making part of it, same applies to llm as a judge and evals too.
Best use case of this mode is fast and cheap computer use, of using a complex application like eg. excel, photo shop, blender etc for longer horizon work, so major application or use cases are all vision dependent.
Another possible future that can be imagined with, system one generation is with on the fly UI and GUI code gen in this way and also real time video/env generation like with many real time 4d models.
What's great or interesting with jev, is its speed and cost and direct integration with code or s/w, not its capabilities, so it might not be ready yet for replacing current models in any way.
On Jev & System One Thinking Models.
I don't know if it's (jev) a single shot structured generation as parallel gen (maybe diffusion llm) via a smaller sized model that is leading to such numbers (speed and cost), but it's a good way to explain system one thinking modes of AI models we might need or see in future.
What we see with examples or use cases of jev is a good example to explain system one thinking modes of llm's, these are super quick, cheap and lower intelligence outputs or actions done by AI.
These have great implications for agentic tool calling, orchestration and computer use, as those will become this fast and this cheap in future, same applies for physical AI (robotic/physical actions and inference) like FSD's (full self driving systems) via system one thinking mode of AI models (llm's) that can work in parallel to normal thinking or other processing.
Great example of this are interaction based duplex models that can do both voice interactions and thinking in parallel, simultaneously super fast (in real / speech time) i.e real time speech and thinking, now imagine that like jev with agentic tool use, computer (also system and env) use and physical AI (physical actions and use cases).
With decision making of higher cognitive and intelligence requirement, current AI tech can't do even frontier level intelligent decision making at system one thinking level, but in future that might be possible, so that is wrt to decision making part of it, same applies to llm as a judge and evals too.
Best use case of this mode is fast and cheap computer use, of using a complex application like eg. excel, photo shop, blender etc for longer horizon work, so major application or use cases are all vision dependent.
Another possible future that can be imagined with, system one generation is with on the fly UI and GUI code gen in this way and also real time video/env generation like with many real time 4d models.
What's great or interesting with jev, is its speed and cost and direct integration with code or s/w, not its capabilities, so it might not be ready yet for replacing current models in any way.
Ditesh retweeted
Through the VOICE trial, Terry is using his Neuralink implant to help fine-tune a brain-to-voice interface for himself and others who can’t speak. He trained the algorithm first by miming speech as best he could, then by simply thinking the words and hearing them come out in his own natural voice.
Powered by Grok Voice from @SpaceXAI
So what I get at is, Build a durable capability/technology, and use a narrowly chosen product/service as the instrument through which you train, test, finance and compound that capability/technology or use growth, investments for same till something else can substitute for them.
Capability/Technology vs Product paradox
I'm quite conflicted with the product aspect of AI venture, and reason is not that I couldn't think of anything to build, but most current things to build won't last or make it in the long run, the one's surviving will transform into something else, as AI itself is evolving and hasn't reached it's final form.
So my reason to be conflicted is what's expected at smaller levels is to build something that makes sense in building in current AI gold rush, for example agents and stuff related to them, they might even generate revenue, but they are not durable and viable in long run i.e even maybe 3 years, so most might crash and burn or transform into something else.
I'm not pessimistic it's just how these things are, building capability/technology is a solution that then can be realized as some product in future, but then vc's wants product trending or having traction in short term, but long term durability is not factored in, it makes no sense, to build something for short term with AI especially when they are not viable economics wise, I would feel being forced into this if done at that scale, to create something that would crash and burn for the sake of increasing adoption or figuring things out.
It's not building a mvp or Ver O of something, then improving it incrementally, it's building something trending for short term, that would get traction but won't be durable, it's like when done at scale, this crash n burn kind of thing, some things could be figured out of them and those winning in it can transform in to things that would make sense eventually.
Maybe reason is it's so darn difficult to figure things out from the start in this rapidly evolving dynamic field, and this is the only way to figure things out, build crash and burn products, but I'm conflicted by this and this doesn't seem right fit for me or my research or capabilities, especially when I know about that, then doing so would be hypocrisy.
Of course this doesn't applying to trying/testing things out, or ideating, but this is the reason people feel AI is forced and slop and has no value to human society, because that long term durable value is not what's being focused on or pushed right now, it's kinda sad, technology wise things are looking great, products wise I'm not sure, everything seems experimental and work in progress maybe it's still early days and adoption is the main focus right now, to make AI a success.
Capability/Technology vs Product paradox
I'm quite conflicted with the product aspect of AI venture, and reason is not that I couldn't think of anything to build, but most current things to build won't last or make it in the long run, the one's surviving will transform into something else, as AI itself is evolving and hasn't reached it's final form.
So my reason to be conflicted is what's expected at smaller levels is to build something that makes sense in building in current AI gold rush, for example agents and stuff related to them, they might even generate revenue, but they are not durable and viable in long run i.e even maybe 3 years, so most might crash and burn or transform into something else.
I'm not pessimistic it's just how these things are, building capability/technology is a solution that then can be realized as some product in future, but then vc's wants product trending or having traction in short term, but long term durability is not factored in, it makes no sense, to build something for short term with AI especially when they are not viable economics wise, I would feel being forced into this if done at that scale, to create something that would crash and burn for the sake of increasing adoption or figuring things out.
It's not building a mvp or Ver O of something, then improving it incrementally, it's building something trending for short term, that would get traction but won't be durable, it's like when done at scale, this crash n burn kind of thing, some things could be figured out of them and those winning in it can transform in to things that would make sense eventually.
Maybe reason is it's so darn difficult to figure things out from the start in this rapidly evolving dynamic field, and this is the only way to figure things out, build crash and burn products, but I'm conflicted by this and this doesn't seem right fit for me or my research or capabilities, especially when I know about that, then doing so would be hypocrisy.
Of course this doesn't applying to trying/testing things out, or ideating, but this is the reason people feel AI is forced and slop and has no value to human society, because that long term durable value is not what's being focused on or pushed right now, it's kinda sad, technology wise things are looking great, products wise I'm not sure, everything seems experimental and work in progress maybe it's still early days and adoption is the main focus right now, to make AI a success.