5 ms·

I can’t speak for other startups, but I applied to the most recent YC batch with my idea for making AI proactive instead of reactive, and pre-being selected I’v

I can’t speak for other startups, but I applied to the most recent YC batch with my idea for making AI proactive instead of reactive, and pre-being selected I’ve published a paper on recursive self-improvement mapped to the Epoch AI data. I contacted a professor from a university in the UK and he responded since he was working on similar work, then asked me if I wanted to meet with him. We talked for about an hour since we had overlapping results and different methods, specifically different assumptions. I say all that to say, as a physics student getting my undergrad, simply doing independent research and speaking to experts about it enabled me to network with someone I otherwise likely wouldn’t know. For young people getting into any business, research is a great way to meet new people.

threethirtytwo 2mo ago
Recursive self improvement is the beginning of the end.
- himata4113 2mo ago
  it's actually the end of the beginning, there's walls that self-improving models hit that are never overcome even when given vast amounts of time.
  - threethirtytwo 2mo ago
    I don't think anyone has ever tried having a model fully autonomously train a model that is better than it.
    - himata4113 2mo ago
      Everyone including myself has attempted this relentlessly, it just doesn't work out beyond some arbitrary improvement in specific tested categories instead of broader capability increase.
      - dgellow 2mo ago
        Did you publish on that topic? Just curious, would love reading more
        himata4113 2mo ago
        There's a pretty large paper on the topic: https://arxiv.org/pdf/2606.15497 https://arxiv.org/pdf/2606.15497, but results are somewhat mixed.
        threethirtytwo 2mo ago
        [flagged]
        threethirtytwo 2mo ago
        My response, which was flagged by automation (I used AI to verify the paper and it says the paper is NOT what you claim it is), was spot on. This paper is not recursive self improvement, it is AI in a feedback loop, but it is not improving itself. It turns your entire premise into something questionable. You said this one done everywhere by many people including you. Then you present a paper which WAS NOT recursive self improvement so it clearly is not "everywhere". To be real with you, I don't think you're being entirely honest. You're not presenting a fair argument. You have bias. Recursive self improvement can only be done by frontier labs who created these models in the first place. Academia simply Does not have the funds. If this is being done, we're not privvy to it. The amount of compute to build these things is astronomical.
    - DonHopkins 2mo ago
      How does not thinking advance your argument? Have you tried asking people instead of not thinking?
- skeptic_ai 2mo ago
  ChatGPT 5.6 Sol ultra still can get basic things written cleanly. I still need to tweak it quite a few times to get it right. Super useful but if you vibe code you’re in a huge mess after 1 billion tokens.
- Windchaser 2mo ago
  Recursive self-improvement only takes off when the improvements are large enough. We could set an LLM loose on itself now, self-improving its own code, but it's going to only make small improvements, and those will quickly peter out. If you think of it in terms of calculus, the sum of the improvements converges to a finite value, or at least the rate of improvements drop with time. And even if we build a much better AI capable of performing substantial self-improvement, a burst of such improvement might peter out quickly. Maybe the AI makes fundamental breakthroughs in computing technology once, and then even with its newly improved capabilities, the next breakthrough is smaller, and the one after that is smaller still. Or maybe we hit a period of S-shaped growth, which seems super-exponential at first but then tapers off. It's wise to be a bit skeptical in our dreams about the future. Yes, society might transform quite radically quite quickly, or on the other hand, maybe the "singularity" isn't even possible. We don't yet know what the scientific limits are.
sillysaurusx 2mo ago
That's also a reason the big labs stopped. Publishing is most valuable to people who have no other way to get the attention of smart strangers. Once you can hire nearly anyone and everyone already returns your calls, the main remaining effect of publishing is to tell your competitors which things worked. This is what happens to every field as it turns from a science into an industry. Chemists published freely until dyes started being worth money, and then the interesting work moved into company labs and stopped coming out.
- teleforce 2mo ago
  Which is a very ironic and selfish situation when your business model dependent mostly on model training based on available published data, academic and non-academic.
  - ReactiveJelly 2mo ago
    That's why we should enforce copyleft
    - fluoridation 2mo ago
      That has nothing to do with anything. If you publish a copyleft paper, that doesn't compel someone who makes a product based on your paper to publish more papers.
      - fc417fc802 2mo ago
        The GNU RPL (research public license), a viral knowledge license. By reading this paper you are legally obligated to openly publish all vaguely related future research that you perform.
        fluoridation 2mo ago
        There's already an RPL, incidentally: https://en.wikipedia.org/wiki/Reciprocal_Public_License https://en.wikipedia.org/wiki/Reciprocal_Public_License Your RPL wouldn't be enforceable. Copyright doesn't deal with abstract ideas passing through people's minds. Even the GPL is kind of in a gray area because the virality feature and its definition of "derivative work" have never been tested in court, to my knowledge. Maybe under contract law, no idea. If nothing else, I'd love to hear a verdict.
colingauvin 2mo ago
Curious what you mean by proactive? Could you share a bit more?
- ninjahawk1 2mo ago
  Happily, current AI is interacted with in a reactive loop. I open the Claude app, CLI, whatever, say my prompt, get an output. I personally wanted an AI that was able to reach out to me about my life before I had to reach out to it. An example, a friend just emailed me asking to meet for at 1pm but I have class at 1:30, so a proactive AI would see that conflict and send me a notification about it, asking if the proposed email it drafted works, then I press send. My personal setup tracks my mouse movement, keyboard, what’s on my screen, and keeps track of what I’m working on through files on my PC. It can update the backend and then restart it on it’s own, meaning I can develop the thing itself while being away from my PC. The capabilities are more than what I’ve listed, but I want to avoid being too preachy about something I made. Here’s the repo if you wanted to take a look, it’s open-source and connects to the iPhone app: https://github.com/getorb/Orb-Backend https://github.com/getorb/Orb-Backend
  - mikepurvis 2mo ago
    I'm very interested in this kind of thing as a kind of ADHD brain augment, like it's monitoring my slack, github, email, calendar, active terminals, etc, and helps me prioritize what I should work on as well as weighing whether this or that ping is worth interrupting me for. I assumed that's what openclaw basically was, but is Orb different from that? And is it fundamentally a different model from the request/response, or is it just request/response in an autonomous loop?
    - ninjahawk1 2mo ago
      Exactly what my thoughts were when I first heard about Openclaw, that’s the exact idea of Orb that you pointed out. Letting you make less decisions, right now AI gives you answers but still requires decisions based on the outputs it gives you. This would deepen the actual ability of agents in those channels you listed. On a fundamental level the backend was designed to do as little LLM calls as possible, for instance it’ll do scans of my screen every 15 seconds, log what’s on it and what’s going on, and store it in a local database, then Orb reviews the entire database every 6 hours for me. Then it’ll schedule wakeups for itself throughout the day, up to 4 so it doesn’t waste my tokens, and schedule notifications based on the last database dump it made. I have my Claude Code, Codex, and Grok Build all useable by using the “claude -p; codex -p…etc” so you can also use multiple CLI’s in conjunction at the same time on different projects or the same project. So your question about a loop is kind of right, but it really just collects your data all day and stores it locally on your PC then calls the LLM of your choice and it reviews all the data and makes those proactive moves we’ve discussed. You could theoretically get it to always be scanning by an LLM but that would be a drastic waste of money from what I’ve seen since most things don’t require a call.
chrisweekly 2mo ago
Yes! And, whether it stems from research or not, putting yourself out there and talking to other people is one of the most fundamentally important things you can do for your career and your personal development as a human being.
- johnnyanmac 2mo ago
  Tell that to my tinder profile. I jest, no tinder. But still, be ready for a lot of non-responses if you're not actively in college. It can feel like a lonely world out there despite theoretically being hundreds of potential people you'd be able to talk to for hours.
  - nicbou 2mo ago
    I think that accepting rejection is part of the being extroverted and meeting strangers. It’s better than waiting for the perfect opening or only talking to people you are sure will be receptive. Or so I was told.
BetterThanSober 2mo ago
LinkedIn can and should be a platform like this, for professionals and catered to professionals. It is a shame that it's now a cesspool of engagement bait and larp entrepreneurs
- epolanski 2mo ago
  LinkedIn is not and should not be a medium for scientific discussions.
  - BetterThanSober 2mo ago
    Agreed, not discussions but networking
Arshad-Talpur 2mo ago
I agree with your point of view we need more research papers specially Breakthroughs AI is acheiving must be published and verified with sound mathematical backings, At our startup we are also moving in this direction perhaps I have already onboarded an Applied Math Phd from Germany to help me in researching.
MassiveOwl 2mo ago
What exactly makes it proactive? I created a similar system that ran on a ticker, but then could also set itself to run in n seconds in the future. RSI came from the actions available to the system and granting the ability to modify itself. The prompt was a string of messages made from static and dynamically generated sections (like memories or plans of tasks or outputs from actions taken on the previous turn). It worked really well
brandonb 2mo ago
FWIW, several YC startups have also published ML research--there were several of us at the last NeurIPS. So perhaps this is a niche that startups can occupy if the big labs don't.

threethirtytwo 2mo ago

Recursive self improvement is the beginning of the end.

himata4113 2mo ago
it's actually the end of the beginning, there's walls that self-improving models hit that are never overcome even when given vast amounts of time.
- threethirtytwo 2mo ago
  I don't think anyone has ever tried having a model fully autonomously train a model that is better than it.
  - himata4113 2mo ago
    Everyone including myself has attempted this relentlessly, it just doesn't work out beyond some arbitrary improvement in specific tested categories instead of broader capability increase.
    - dgellow 2mo ago
      Did you publish on that topic? Just curious, would love reading more
      - himata4113 2mo ago
        There's a pretty large paper on the topic: https://arxiv.org/pdf/2606.15497 https://arxiv.org/pdf/2606.15497, but results are somewhat mixed.
        threethirtytwo 2mo ago
        [flagged]
        threethirtytwo 2mo ago
        My response, which was flagged by automation (I used AI to verify the paper and it says the paper is NOT what you claim it is), was spot on. This paper is not recursive self improvement, it is AI in a feedback loop, but it is not improving itself. It turns your entire premise into something questionable. You said this one done everywhere by many people including you. Then you present a paper which WAS NOT recursive self improvement so it clearly is not "everywhere". To be real with you, I don't think you're being entirely honest. You're not presenting a fair argument. You have bias. Recursive self improvement can only be done by frontier labs who created these models in the first place. Academia simply Does not have the funds. If this is being done, we're not privvy to it. The amount of compute to build these things is astronomical.
  - DonHopkins 2mo ago
    How does not thinking advance your argument? Have you tried asking people instead of not thinking?
skeptic_ai 2mo ago
ChatGPT 5.6 Sol ultra still can get basic things written cleanly. I still need to tweak it quite a few times to get it right. Super useful but if you vibe code you’re in a huge mess after 1 billion tokens.
Windchaser 2mo ago
Recursive self-improvement only takes off when the improvements are large enough. We could set an LLM loose on itself now, self-improving its own code, but it's going to only make small improvements, and those will quickly peter out. If you think of it in terms of calculus, the sum of the improvements converges to a finite value, or at least the rate of improvements drop with time. And even if we build a much better AI capable of performing substantial self-improvement, a burst of such improvement might peter out quickly. Maybe the AI makes fundamental breakthroughs in computing technology once, and then even with its newly improved capabilities, the next breakthrough is smaller, and the one after that is smaller still. Or maybe we hit a period of S-shaped growth, which seems super-exponential at first but then tapers off. It's wise to be a bit skeptical in our dreams about the future. Yes, society might transform quite radically quite quickly, or on the other hand, maybe the "singularity" isn't even possible. We don't yet know what the scientific limits are.

sillysaurusx 2mo ago

That's also a reason the big labs stopped. Publishing is most valuable to people who have no other way to get the attention of smart strangers. Once you can hire nearly anyone and everyone already returns your calls, the main remaining effect of publishing is to tell your competitors which things worked. This is what happens to every field as it turns from a science into an industry. Chemists published freely until dyes started being worth money, and then the interesting work moved into company labs and stopped coming out.

teleforce 2mo ago
Which is a very ironic and selfish situation when your business model dependent mostly on model training based on available published data, academic and non-academic.
- ReactiveJelly 2mo ago
  That's why we should enforce copyleft
  - fluoridation 2mo ago
    That has nothing to do with anything. If you publish a copyleft paper, that doesn't compel someone who makes a product based on your paper to publish more papers.
    - fc417fc802 2mo ago
      The GNU RPL (research public license), a viral knowledge license. By reading this paper you are legally obligated to openly publish all vaguely related future research that you perform.
      - fluoridation 2mo ago
        There's already an RPL, incidentally: https://en.wikipedia.org/wiki/Reciprocal_Public_License https://en.wikipedia.org/wiki/Reciprocal_Public_License Your RPL wouldn't be enforceable. Copyright doesn't deal with abstract ideas passing through people's minds. Even the GPL is kind of in a gray area because the virality feature and its definition of "derivative work" have never been tested in court, to my knowledge. Maybe under contract law, no idea. If nothing else, I'd love to hear a verdict.

colingauvin 2mo ago

Curious what you mean by proactive? Could you share a bit more?

ninjahawk1 2mo ago
Happily, current AI is interacted with in a reactive loop. I open the Claude app, CLI, whatever, say my prompt, get an output. I personally wanted an AI that was able to reach out to me about my life before I had to reach out to it. An example, a friend just emailed me asking to meet for at 1pm but I have class at 1:30, so a proactive AI would see that conflict and send me a notification about it, asking if the proposed email it drafted works, then I press send. My personal setup tracks my mouse movement, keyboard, what’s on my screen, and keeps track of what I’m working on through files on my PC. It can update the backend and then restart it on it’s own, meaning I can develop the thing itself while being away from my PC. The capabilities are more than what I’ve listed, but I want to avoid being too preachy about something I made. Here’s the repo if you wanted to take a look, it’s open-source and connects to the iPhone app: https://github.com/getorb/Orb-Backend https://github.com/getorb/Orb-Backend
- mikepurvis 2mo ago
  I'm very interested in this kind of thing as a kind of ADHD brain augment, like it's monitoring my slack, github, email, calendar, active terminals, etc, and helps me prioritize what I should work on as well as weighing whether this or that ping is worth interrupting me for. I assumed that's what openclaw basically was, but is Orb different from that? And is it fundamentally a different model from the request/response, or is it just request/response in an autonomous loop?
  - ninjahawk1 2mo ago
    Exactly what my thoughts were when I first heard about Openclaw, that’s the exact idea of Orb that you pointed out. Letting you make less decisions, right now AI gives you answers but still requires decisions based on the outputs it gives you. This would deepen the actual ability of agents in those channels you listed. On a fundamental level the backend was designed to do as little LLM calls as possible, for instance it’ll do scans of my screen every 15 seconds, log what’s on it and what’s going on, and store it in a local database, then Orb reviews the entire database every 6 hours for me. Then it’ll schedule wakeups for itself throughout the day, up to 4 so it doesn’t waste my tokens, and schedule notifications based on the last database dump it made. I have my Claude Code, Codex, and Grok Build all useable by using the “claude -p; codex -p…etc” so you can also use multiple CLI’s in conjunction at the same time on different projects or the same project. So your question about a loop is kind of right, but it really just collects your data all day and stores it locally on your PC then calls the LLM of your choice and it reviews all the data and makes those proactive moves we’ve discussed. You could theoretically get it to always be scanning by an LLM but that would be a drastic waste of money from what I’ve seen since most things don’t require a call.

chrisweekly 2mo ago

Yes! And, whether it stems from research or not, putting yourself out there and talking to other people is one of the most fundamentally important things you can do for your career and your personal development as a human being.

johnnyanmac 2mo ago
Tell that to my tinder profile. I jest, no tinder. But still, be ready for a lot of non-responses if you're not actively in college. It can feel like a lonely world out there despite theoretically being hundreds of potential people you'd be able to talk to for hours.
- nicbou 2mo ago
  I think that accepting rejection is part of the being extroverted and meeting strangers. It’s better than waiting for the perfect opening or only talking to people you are sure will be receptive. Or so I was told.

BetterThanSober 2mo ago

LinkedIn can and should be a platform like this, for professionals and catered to professionals. It is a shame that it's now a cesspool of engagement bait and larp entrepreneurs

epolanski 2mo ago
LinkedIn is not and should not be a medium for scientific discussions.
- BetterThanSober 2mo ago
  Agreed, not discussions but networking

Arshad-Talpur 2mo ago

I agree with your point of view we need more research papers specially Breakthroughs AI is acheiving must be published and verified with sound mathematical backings, At our startup we are also moving in this direction perhaps I have already onboarded an Applied Math Phd from Germany to help me in researching.

MassiveOwl 2mo ago

What exactly makes it proactive? I created a similar system that ran on a ticker, but then could also set itself to run in n seconds in the future. RSI came from the actions available to the system and granting the ability to modify itself. The prompt was a string of messages made from static and dynamically generated sections (like memories or plans of tasks or outputs from actions taken on the previous turn). It worked really well

brandonb 2mo ago

FWIW, several YC startups have also published ML research--there were several of us at the last NeurIPS. So perhaps this is a niche that startups can occupy if the big labs don't.