yunyu

When you speak or type, you speak one word at a time. When you move, you actuate one muscle at a time.

Does this mean that a single incorrect word or twitch will completely derail the task you’re trying to performance? Or will you, like any other intelligent being, recognize it and compensate?

It's pretty clear reading from these comments that most HN members have a 2023-era impression of LLMs.

Modern chain-of-thought models with RL post training on verifiable tasks + realistic environments + rubrics are worlds apart from models trained on a simple next token prediction objective.

More money goes into the rubrics and RL environments than individual training runs themselves.

(Yes, at inference-time LLMs still output words one at a time, much like human speakers. But don't confuse the mechanism with the training objective.)

On the flip side, lightning puts the part that wears out most quickly (the retention springs) inside the device, while USB-C does the opposite. The same argument can be made for both sides.

Mercor has around 5 customers that make up 95% of its revenue. Anybody who needs to know about them already does.

What do you guys do differently than Profound or Airops?

This has been happening for the past year on verifiable problems (did the change you made in your codebase work end-to-end, does this mathematical expression validate, did I win this chess match, etc...). The bulk of data, RL environment, and inference spend right now is on coding agents (or broadly speaking, tool use agents that can make their own tools).

Recent advances in mathematical/physics research have all been with coding agents making their own "tools" by writing programs: https://openai.com/index/new-result-theoretical-physics/

You can just use it outside of the API?

I was trying to point out the selection bias in your experiences, since your queries are probably lower-intent than for the typical consumer. Good discussion though.

If you search for low-value consumer queries or aren't a buyer for high-ticket B2B items, you will get less relevant ads. Your personal experiences don't extrapolate well here: advertising for high-value goods is totally different from advertising for low-value goods.

Example: hypertargeted ads for F-35 engine upgrades in the DC metro - https://x.com/JosephPolitano/status/1683476652276236295

Is that clear enough?

...that was my point? You are so close to getting it.

You are talking about consumer marketing. I am talking about B2B marketing for prescription medications, enterprise SaaS, etc. These are separate markets and the analogies don't quite hold here - the scam problem etc is practically nonexistent for high-LTV goods with high bid costs, and newcomers are typically well funded enough to periodically outbid incumbents (or implement better targeting). The big-ticket B2B products that one hears of from word-of-mouth are usually the worse ones, since there is rarely any "going viral" to speak of.

> That the engine's website puts a different thing that whatever the search algorithm thought was the best thing at the top because an advertiser paid for it to be so, is strictly worse.

This is not clearly worse than the result being selected by the whims of some arbitrary Google engineer, or being easily gamed by SEO blogspam bots. At least the advertiser stands to lose something if they bid incorrectly.

>How long had ChatGPT been out before OpenAI's first ad for it?

Just because some products were able to grow organically doesn't imply that paid marketing never benefits startups. This is a false equivalence.

I also find it funny that the vast majority of your example products (everything except Huel or Aeropress?) make a lot of money from advertising. Maybe consider why they still exist.

Yes, advertising is more relevant and works better for people with power over large purchasing decisions as the bidders have more at stake. Maybe you aren't in the market for plumbers or running shoes, and are instead looking for "download vlc media player". This doesn't contradict anything I said?

My favorite form of definitely-not-advertising :)

Your value to advertisers is probably less than 1% of that of a single doctor or corporate VP. It makes sense that your queries are lower intent - this is hardly contradictory. Fortunately, you are an edge case wrt how firms are actually spending their money, so we’ll leave it at that.

> Medicine has a pretty good system for getting knowledge out to doctors as far as I can tell.

Yes, it does - it’s called advertising. In the US, the average promotional spend per physician exceeds $20k/yr. As a result, a lot more patients are able to quickly benefit from new medications like Dupixent or Ozempic as a result of wider awareness.

Suppose we banned Google ads and you are searching for a plumber. You are now entirely at the whims of whoever designs the ranking algorithm on Google/the Yellow Pages, who has nothing at stake here. Meanwhile, advertisers have to bid for your attention - making them at least somewhat aligned with your buying intent.

The same applies for doctors searching for state of the art diabetes treatments. It’s hard to say that relying on a fuzzy notion of “legitimacy” (or entrenched status-quo cliques) is a more fair system.

Got it. So you want attention to be controlled by the whims of academic/government/publishing bureaucrats or black-box ranking algorithms who are the arbitrators of legitimacy. I can't say I agree with that opinion, but different strokes for different folks.

The fact that your Facebook ads are worse is probably because you're in the EU. I'm in the US, and I am getting fairly relevant ads for Broadway shows, data infrastructure products, discounted hotel packages, and climbing gym subscriptions - things that I am actually considering purchasing. And we haven't even brought up intent-based ads (Google search).

Word of mouth benefits incumbents. Advertising at least enables newcomers to temporarily burn money to gain mindshare, while “slow diffusion” will lock society into a “nobody ever got fired for buying IBM” state forever.

What are legitimate sources in your definition? Should physicians be expected to spend all their free time reading every single study in every medical journal or conference, even for niche areas that they don't usually encounter? Should the average diabetic/arthritic patient need to obsessively pore over academic reports to stay informed about their condition? Should advertisers be banned from sponsoring journals or conferences? This is an extremely ill informed line of reasoning.

Advertising enables innovation-producing firms to drive awareness of their services in a cost effective manner, and for less informed consumers to understand what is available on the market. Your typical physician might not be fully caught up on what is the state of the art in arthritis treatments, but advertising enables this to happen.

>A conscious textual agent would need something like a unified narrative environment with real feedback: symbols that maintain identity over time, a stable substrate where “being someone” is definable, the ability to form and test a hypothesis, and experience the consequences. LLMs don’t have that. They exist in a shifting cloud of possibilities with no single consistent reality to anchor self-maintaining loops. They can generate pockets of local coherence, but they can’t accumulate global coherence across time.

These exist? Companies are making billions of dollars selling persistent environments to the labs. Huge amounts of inference dollars are going into coding agents which live in persistent environments with internal dynamics. LLMs definitely can live in a world, and what this world is and whether it's persistent lie outside the LLM.

We do? Tool use started coming in vogue around 2023

Great idea. I always end up having to tag the relevant files/abstractions anyways to avoid having the LLM produce duplicated slop, and something like this makes collecting this info much easier.

It’s good for collecting a lot of computer use trajectory data

The goal of most companies is to maximize returns to shareholders, not society or employees. If it wasn't that I wouldn't have invested.

As a shareholder of many companies, I would be disappointed if the management team made decisions that disproportionately benefited employees at the expense of other non-employee shareholders.

There is a lot of pretraining data available around screen recordings and mouse movements (Loom, YouTube, etc). There is much less pretraining data available around navigating accessibility trees or DOM structures. Many use cases may also need to be image aware (document scan parsing, looking at images), and keyboard/video/mouse-based models generalize to more applicants.

The burger flippers (and society at large)

It is better to empower 1 Noam Shazeer or Terence Tao than to prop up 10 million burger flippers.

I wonder what ethnicity the majority of Taiwanese are, where they came from, and how recently...

I wonder what the ethnicity of the Americans who founded, run, and developed NVIDIA and TSMC are.