Item — HN Reader

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.

8 comments

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.

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.

That's why we should enforce copyleft

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.

I had some meeting with non-creative business person, just chatting over coffee. I wondered how much he used from our meeting for his enrichment.

Public papers with described research give advance to similar people. They jump over you and in many cases they give nothing back.

(Intellectual) greed is everywhere and is cross-border.

In case of public papers there is one extra vulnerability - competitor(s) can build anywhere, under the radar.

That’s why companies are so cautious about publishing their research…

Sure.

But they’re not dependent on my research in particular.

If I don’t publish, what, as a result of my not publishing, happens to the companies dependent on data & research?

Nothing.

i wish more people would publish the things that didn't work. i'd like everything, but the exploration of searched negative space is so wasteful.

Everyone says this in the abstract but to concrete examples they shug and say, of course that approach doesn't work, they did X, Y, Z wrong, they should have given it more effort, it could have worked if done properly / this can obviously never work, everyone knew already, it's nothing new etc.

I think you would be interested in the Journal of Trial and Error: https://journal.trialanderror.org/

> I wish more people would publish the things that didn't work.

This would be a pointless endeavour. One of the most basic mantras of science is "absence of evidence is not evidence of absence". So just because something didn't worked out for you that doesn't mean it doesn't work out for others, or even yourself in the future.

> Chemists published freely until dyes started being worth money

Notably this is exactly what patents are intended to combat. And while US IP law is clearly very broken it does at least largely accomplish this stated goal. Much (but certainly not all) industrial chemistry has made it into the academic literature.

Not that the same logic necessarily applies to AI research (ie algorithms aka math and their implementations). And I'm actually happy about that because the cost of doing the research is so much lower. There's a long list of reasons that the average person living in a residential area can't do industrial chemistry as a hobby.

To your dye example, yttrium indium manganese blue was the first commercially viable inorganic blue pigment discovered in ~200 years, is the only known environmentally safe one, and was openly published in the literature. It's also under an exclusive license. (TBF though unless the chemical is unusually difficult to synthesize not publishing would be rather pointless in this day and age given the utterly absurd capabilities of modern analytical techniques.)

It depends on the field, and to some extent on the particular patent officer assigned, but one of the failings of the US patent system is many patents don't actually provide enough details to reproduce the result. And even if they do they don't (usually) include ancillary information like the discovery process, properties of the invention etc.

>> To your dye example, yttrium indium manganese blue was the first commercially viable inorganic blue pigment discovered in ~200 years, is the only known environmentally safe one, and was openly published in the literature. It's also under an exclusive license.

Gee, I wonder what was wrong with the previous blue pigments and why it was so important to have this one under an exclusive license.

Cobalt blue is a blue pigment made by sintering cobalt(II) oxide with aluminium(III) oxide (alumina) at 1200 °C. Chemically, cobalt blue pigment is cobalt(II) oxide-aluminium oxide, or cobalt(II) aluminate, CoAl2O4. Cobalt blue is lighter and less intense than the (iron-cyanide based) pigment Prussian blue.

https://en.wikipedia.org/wiki/Cobalt_blue

Oh right.

P.S. Don't lick your brushes.

Hot take, copyright law should be reformed to be more like patents.

Want the government / courts to stop your employees leaking source code? Escrow the code, and release it in 20 years.

The residuals on 20 year code is so close to zero that the costs vs benefits of longer IP protection is not in the public interest.

The startup landscape has also changed noticeable compared to 5 or 10 years ago. A team of smart/credible people could get funding for an idea and build a product + publish, knowing there was a six month lead time for anyone to copy them and ship. These days, the barrier to ship code is zero. People can copy your business over a weekend, so there is much more urgency to establish product market fit and build a “moat”. Publishing timelines are now at odds with the pace of go-to-market and VC funding timelines.

I think in theory this is true but in practice I don't see loads more good apps or products. There's a paradox here I think. I just don't see loads of quality competitors popping up I actually think it makes building something harder because the barrier to entry just gets higher somehow.

Over the weekend, I analyzed public YouTube tutorials of 4 SaaS products in the same domain. Then I asked the coding agent to define an API where they converge — prior art. Today I'm working on creating dashboards that won't touch anyone's copyright or IP.

A couple of days ago, something interesting happened. The agent works in an iteration loop where each iteration is an endpoint. I let it run overnight. In the morning it was still cranking away even though it had finished all the API endpoints. It had found a changelog from one of the companies listing every single feature and bug, and decided on its own to implement every item as an iteration.

We are 2 to 3 months away from coding agents replicating solving the edge cases of most SaaS applications.

There are other reasons. One is for career development of your researchers. Another is to be good citizens, in good standing, in the community of scientific researchers. Another is to flex on people and buy class and respectability - to comport oneself as the “old money” does.

Publishing is an act of spreading science.

The biggest incentive besides the purpose is really fame and tangible results for your academic career.

Most AI labs, especially the top private ones, don't have particular incentives to publish their results and findings.

On the contrary, I've found that stuff in the mining industry is actually well published, to the point of replicability (one of our projects is essentially copying the patented design of a well known mining technology company, albeit with some modifications, as the patent will expire next month).

Perhaps because the real moat in the industry isn't the technology per se, but the concessions.

It’s funny because no frontier labs would exist if the Google transformer paper wasn’t published

>> Chemists published freely until dyes started being worth money, and then the interesting work moved into company labs and stopped coming out.

I think this also makes the case for Industrial Espionage

Such a biased and utilitarian view. The goal of publishing is to communicate your findings to the scientific community for the sake of advancing our knowledge as a society.

Of course you might want to keep _some_ of this knowledge as trade secret, but then don’t claim you are doing research.

You are advancing your shareholder’s interests, not that of the broader society.

The profit motive is the reason that most of the things that are shit, are shit.

This academic publish and industry doesn't is an incorrect view. Academics also hold off from publishing groundbreaking science in order to protect their research edge. You shouldn't see this as industry bad as driven by profits and academia is free from career incentives and money.

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

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.

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.

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.

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.

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.

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

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

> 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.

I don't mean to downplay your work, but I think you should come up with a better example use case. Automating away interactions with friends is pretty much the last thing I want AI to do.

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?

Hmm I might be completely missing something here (I’m not a machine learning person), but how is that “pro-active”? The mode is still taking in an input (or several inputs maybe in this case) and responding to that input, isn’t it? It doesn’t seem much different than current capabilities of the various agentic harnesses on the market right now. Again, apologies if I’m missing something obvious here…

OK really cool. I've been working on something similar but I started from the opposite direction. I first developed a database that pulls in data from as many personal sources as possible, and then runs it through a gradual annotation funnel going from coarse to fine, culminating in embedding for semantic and lexical search. At that point, it computes timelines and has resolution gradients along timelines (older = lower resolution, newer = higher resolution. Then I built an MCP server for that which can be used to query it. What's missing is the afferent arm, so these might be a really nice pairing.

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

> Happily, current AI is interacted with in a reactive loop. I open the Claude app, CLI, whatever, say my prompt, get an output.

"Current AI" is not limited to LLM offerings. There are many AI algorithms which can assist in what you specify thusly:

> I personally wanted an AI that was able to reach out to me about my life before I had to reach out to it.

Consider a forward chaining inference engine ("expert system") provided with relevant asynchronous percepts from the deployed environment to reason about. This could serve as an initiator of a "proactive AI".

But it's not proactive then? There's just a hidden loop or some cron-like signal that feeds data to a reactive loop..

I'd imagine proactive as something like, hmm, no signal from X, I wonder how they're doing...

What’s so innovative to send an auto message every x time or based on events: “wake up and check if you got anything to do”.

“Event x happened at y time”

How does this differ from Hermes / Openclaw / Vellum?

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

LinkedIn is not and should not be a medium for scientific discussions.

Agreed, not discussions but networking

Recursive self improvement is the beginning of the end.

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.

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.

I don't think anyone has ever tried having a model fully autonomously train a model that is better than it.

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.