nlpnerd

This is why every other crypto company is now an AI company.

This is a fascinating example of perhaps why a move towards a "world model" or some better form of representation can be helpful.

Honestly, the article title is a bit of clickbait. The main complaint is about people on medical leave being disadvantaged by the token usage as performance metric that was introduced.

"Nadella argues that if AI companies get to freely scrape the internet to train their models, it’s only fair that enterprises get to study — or “distill” — those models in return."

Two wrongs don't make one right

AI coding forces people to frontload a lot of the detailed thinking that used to take place over an epic or sprint. It's not a very natural way of working, and in fact is quite contrary to the iterative development style that most devs are already used to.

I think a lot of ppl assume that usage here is simply for AI coding, which is easily governed. I suspect that the more tricky issue are those usage powering workflows and application logic that cannot be easily throttled down.

Yeah

That's kind of stretching things. It's not that software cannot be maintained but the cognitive load of building and maintaining software seems to be now backloaded to the maintaining part and people need to be aware of this.

Well...we didn't sack anybody if you were wondering. I feel like Ford's issues is mainly overestimating AI's capabilities. Whereas our case is more of a need for re-engineering processes to be more AI-native.

That is an idealistic take without business sense. Startups (and individual hackers in this case) exists to take this kind of radical bets because the risk/reward profile is asymmetrically in their favour. Whereas for an enterprise, the risk/reward is inverse.

If Peter Steinberger is able to generate even a 100M this year from Clawdbot what he has is a multi billion dollar business that would be life-changing even for a successful entrepreneur like him who is already a multi-millionaire. If it collapses from the security flaws, and other potential safety issues he loses nothing, starting from zero and going back to it. Peter Steinberger (and startups in general) have a lot to gain and very little or close to nothing to lose.

The iPhone generated 400B in revenue for Apple in 2025. Clawdbot even if it contributes 4B in revenue this very year would not move the needle much for Apple. On the contrary, if Apple rushes and botches releasing something like this they might just collapse this 400B/annum income stream. Apple and other large enterprises (and their execs) have a lot to lose and very little to gain from rushing into something like this.

"Paper after paper shows these things are hiding data, fabricating output, reward hacking, exploiting human psychology, and engaging in other nefarious behaviors best expressed as akin to a human toddler - just with the skills of a political operative, subject matter expert, or professional gambler."

Anthropomorphizing removed, it simply means that we do not yet understand the internal logic of LLM. Much less disturbing than you suggest.

Ex quant researcher in an equity hedge fund.

Yup, doing that.

1) Workflow issues. A chatbot is not the best interface for accomplishing the typical tasks an investor or analyst needs to do. Our workflow are often start around a watchlist/portfolio and news.

2) Most LLM or agentic applications don't give us enough control over the process. Investment research is pretty open-ended and subjective, there is no one absolute approach. Individuals and teams often have some tribal perspective on how to do it or needs certain ground covered to have conviction. Existing solutions do not allow such control.

Definitely makes a lot less economic sense today. Takes easily months of full time work to raise from investors. That level of effort in the early days of a startup could put mean practically putting your pre-MVP/MVP on hold for months.

This is why I always believed that VCs and accelerators that mislead startups about how early they are willing to invest are doing a lot of damage to gullible founders.

Common failure mode is people. Most processes eventually fail when people start to slack on the little things, then the big things. Every part of knowledge management is tedious from capturing things to keeping them organized and discoverable. For any large organizations, there are a ton of knowledge that will remain tribal despite best efforts.

This is why knowledge management is such a popular use for POCs involving LLMs, and ironically also why POCs don't progress into something more permanent

Perplexity just rolled out direct access to the SEC EDGAR database across all their interfaces (Search, Research, and Labs). That means retail investors and researchers now have LLM-friendly access to 10-Ks, 10-Qs, 8-Ks, etc., with real-time Q&A and summarization layered on top.

Has anyone tried this out? Curious what the technical implementation might be and what future use cases this unlocks (e.g. automated 10-K comparators, anomaly spotting, regulatory change detection).

Would love to hear thoughts from professionals and serious investors, as well as technologists working in the space.

I have always believed that Chain of Thought basically acts as a form of regularization. LLMs are fundamentally next token predictors without any form or notion of logic, reasoning etc and is as likely (as a probabilistic model) to produce a creative response as something based on facts/principles (or anything that resembles "reasoning").

Asking the LLM to think step by step simply biases it towards the latter. It's still a stochastic parrot but now it sounds logical and that happens to be useful in some cases, regardless of whether we can agree if it's "reasoning".

That is a slight exaggeration, extrapolation on the author's part. What happened was that RL training led to some emergent behavior in R1-Zero (chain-of-thought, and reflection) without being prompted or trained for explicitly. Don't see what is so domain specific about that though.

They don't. This is probably one reason why VCs invest in these companies. There is a natural moat since there is only a very finite number of people in the world has the right experience to raise, and only those who can raise can ever have the experience.

At least until compute cost drop to a cheap enough level...

Not a joke. It's fairly common in South East Asia and even has a name -- White Coffee.

Attitude matters more if we are talking about the longer term. 6 months to a year in the less credential-ed devs will probably be independent. The experienced but uninterested devs are unlikely to have become more interested.

Interesting perspective from the author. Insightful to note that VCs perhaps are not as smart as they would like investors and the public to believe (which I totally agree with, they only human...at least for now) or simply leveraging hype with these recent high profile AI investments.

But I think there are some assumptions within the article worth looking into to better explain why these ventures raised as much as they did.

First of all there is a difference between a "deep-tech" startup trying to develop a better foundation model vs startups trying to build applications (usually a thin wrapper application). The former are the ones raising tens to hundred of millions (apparently as billion dollar valuation or more).

No thin wrapper startups, afaik have raised at the same amount and valuation. These startups on the other hand are actually raising at pretty typical amounts and valuation.

An explanation for this dichotomy is perhaps because there is actually a need for a large upfront investment to get the business of building foundation models off the ground. This by itself of course does not justify giving these startups such high upfront valuation. What may in part justify the valuation is perhaps their moat. The moat which funny enough comes in the form of the high upfront investment required for infrastructure, and perhaps the providing the financial incentives to attract the limited number of star researchers every VC perceive as essential to the success of creating new foundation models. (I assume they are well compensated for taking the leap from top companies already known for top of the market compensation). There are only so many of these star researchers that can raise such monies, which imposes a ceiling on the number of such startups in the near future.

Of course a moat alone is little reason to assign value to a venture. There may be a moat, but is the moat protecting gold or coal? And how much gold justifies a 10x, 20x or even higher multiples? I believe the 10x multiple is simply a rule of thumb based on perceived growth potential and comparison against mature peers in the same segment. Rules of thumb can and should be ignored sometimes, and this may be one of those times. Growth potential looks unprecedentedly good given how fast ChatGPT gained users, and their rumoured revenue growth from 28m in 2022, to a projection of 200m in 2023 and 1B by end of 2024. For reference, a 40% yoy growth counts as hyper-growth and have in the past been used to justify a greater than 20x multiple. The lack of any commonly accepted mature peers based around LLMs, GAI probably only served in these companies' favour given the above results. I am also on the fence as to whether the LLM/model as a service market will be a winner take all or perhaps some kind of competitive oligopoly like we see with cloud computing.

What did you guys do to bring down the latency so much?