Comment by caseyy | Hacker News Reader

Comment by caseyy | Hacker News Reader

Parent Full thread

It's doubtful if there even is a race anymore. The last significant AI advancement in the consumer LLM space was fluent human language synthesis around 2020, with its following assistant/chat interface. Since then, everything has been incremental — larger models, new ways to prompt them, cheaper ways to run them, more human feedback, and gaming evaluations.

The wisest move in the chatbot business might be to wait and see if anyone discovers anything profitable before spending more effort and wasting more money on chat R&D, which includes most agentic stuff. Reliable assistants or something along those lines might be the next big breakthrough (if you ask certain futurologists), but the technology we have seems unsuitable for any provable reliability.

ML can be applied in a thousand ways other than LLMs, and many will positively impact our lives and create their own markets. But OpenAI is not in that business. I think the writing is on the wall, and Sama's vocal fry, "AGI is close," and humanity verification crypto coins are smoke and mirrors.

Saying LLMs have only incrementally improved is like saying my 13 year old has only incrementally approved over the last 5 years. Sure, it's been a set of continuous improvements, but that has taken it from a toy to genuinely insanely useful.

Personally, deep research and o3 have been transformative, taking LLMs from something I have never used to something that I am using daily.

Even if the progress ends up plateauing (which I do not believe will happen in the near term), behaviors are changing; OpenAI is capturing users, and taking them from companies like Google. Google may be able to fight back and win - Gemini 2.5 Pro is great - but any company sitting this out risks being unable to capture users back from Open AI at a later date.

No, it's still just a toy. Until they can make the models actually consistently good at things, they aren't going to be useful. Right now they still BS you far too much to trust them, and because you have to double check their work every time they are worse than no tool at all.

To extend your illustration, 5 years ago no one could train an LLM with the capabilities of a 13 year old human; now many companies can both train LLMs and integrate them into products.

> taken it from a toy to genuinely insanely useful.

Really?

> any company sitting this out risks being unable to capture users back from Open AI at a later date.

Why? I paid for Claude for a while, but with Deepseek, Gemini and the free hits on Mistral, ChatGPT, Claude and Perplexity I'm not sure why I would now. This is anecdotal of course, but I'm very rarely unique in my behaviour. I think the best the subscription companies can hope for is that their subscribers don't realize that Deepseek and Gemini can basically do all you need for free.

>I'm very rarely unique in my behaviour

I cannot stress this enough: if you know what Deepseek, Claude, Mistral, and Perplexity are, you are not a typical consumer.

Arguably, if you have used a single one of those brands you are not a typical consumer.

The vast majority of people have used ChatGPT and nothing else, except maybe clicking on Gemini or Meta AI by accident.

I doubt it. Google is shoving Gemini on everyone’s face through search, and Meta AI is embedded in every Meta product. Heck, instagram created a bot marketplace.

They might not “know” the brand as well as ChatGPT, but the average consumer has definitely been exposed to those at the very least.

DeepSeek also made a lot of noise, to the point that, anecdotally, I’ve seen a lot of people outside of tech using it.

It's been five years. There is no AI killer app. Agentic coding is still hot garbage. Normal people don't want to use AI tools despite them being shoved into every SaaS under the sun. LLMs are most famous among non-tech users for telling you to put glue into pizza. No one has been able to scale their chatbots into something profitable, and no one can put a date on when they'll be profitable.

Why are you still pretending anything is going to come out of this?

I can't square how OpenAi can capture users and presumably retain them when the incumbents have been capturing users for multiple decades and why can they not retain them?

If every major player had an AI option, i'm just not understanding how because OpenAi moved first or got big first, the hugely massively successful companies that did the same thing for multiple decades don't have the same advantage?

Who knows how this will play out, but user behavior is always somewhat sticky and OpenAI now has 400M+ weekly active users. Currently, I'm not sure there is much of a moat, as many would jump if, say, Google released a model that is 10x better. However, there are myriad ways that OpenAI could slowly try to make their userbase even stickier:

1. OpenAI is apparently in the process of building a social network.

2. OpenAI is apparently working with Jonny Ive on some sort of hardware.

3. OpenAI is increasingly working on "memory" as a LLM feature. Users may be less likely to switch as an LLM increasingly feels like a person that knows you, understands you, has a history with you, etc.

4. Google and MSFT are leveraging their existing strengths. Perhaps you will stick with Gemini given deep integration with Android, Google Drive, Sheets, Docs, etc.

5. LLMs, as depressing as this sounds, will increasingly be used for romantic/friend purposes. These users may not want to switch, as it would be like breaking up and finding a new partner.

6. Your chat history, if it can't be easily exported/imported, may be a sticky feature, especially if it can be improved (e.g. easily search, cross-reference, chats, like a supercharged interconnecting note app with brains).

I could list 100 more of these. Perhaps none of the above will happen, but again, they have 400M weekly users and they will find ways to keep them. It's a lot easier to keep users that have a habit of showing up, then getting them in the first place. There's a reason that Google is treating this like an emergency; they are at serious risk of having their search cash cow permanently disrupted if they don't act fast to win.

Very thought provoking reply. #3 sounds the most sticky to me, in the product sense that you'd build "your own LLM/agent" and plug it other services. I heard this on a product podcast [1], think of it like Okta SSO integration: access controls for your personal/sensitive LLM stuff vs all other services trying to get you to use their LLM.

#5 stands out as well as a substantial barrier.

The rest to me our sticky, but no more uniquely sticky than any other service that retains data. Like the switching cost of email or a browser. It does stick but not insurmountable and once the switch is made, it's like why did I wait so long? (I'm a Safari user!)

Anyway, thanks for the thoughtful reply.

[1] https://www.reforge.com/podcast/unsolicited-feedback/the-gre...

6 (can’t export/import chat history) is already a wrap since every user is prohibited from using ChatGPT chat logs to “develop models that compete with OpenAI,” if you export your chats and give it to Gemini or Claude or post it on X and Grok reads it, then you just violated the OpenAI terms, that’s grounds for a permaban or lawsuit for breach of contract (lol) … maybe your companies accept this risk but I’m in malicious compliance mode

Google is alright, but they have similar stupid noncompete vendor lock in rule, and no way to opt out of training, so there’s no real reason to trust Google. Yeah they could ship tool use in reasoning to catch up to o3, but it’ll just be catching up and not passing unless they fix the stupid legal terms.

Claude IDK how to trust, they train on feedback and everything is feedback, and they have the noncompete rule written even more broadly, dumb to use that.

Grok has a noncompete rule but also has a way to opt out of training, so it’s on the same tier of ClosedAI. I use it sometimes for jokey toy image generation crap but there’s no way to use it for anything serious since it has a copypasted closed ai prohibition

Mistral needs better models and simpler legalese, it’s so complicated and impossible to know which of the million legal contracts applies

IMHO meta is the only player, but they shot themselves in the foot by making Llama 4 too big for the local llama community to even use, super dumb, killed their most valuable thing which was the community.

That means the best models we can use for work without needing to worry about a lawsuit, are Qwen, and DeepSeek distills, no American AI is even in the same ballpark, Gemma 3 is refusal king if you even hint at something controversial. basically, America is getting actively stomped by China in AI right now, because their stuff is open and interoperable, and ours is closed and has legal noncompete bullshit, what can we actually build that doesn’t compete with these companies? Nothing

Just to get things right. The big AI LLM hype started end of 2022 with the launch of ChatGPT, DALL-E 2, ....

Most people in society connect AI directly to ChatGPT and hence OpenAI. And there has been a lot of progress in image generation, video generation, ...

So I think your timeline and views are slightly off.

> Just to get things right. The big AI LLM hype started end of 2022 with the launch of ChatGPT, DALL-E 2, ....

GPT-2 was released in 2019, GPT-3 in 2020. I'd say 2020 is significant because that's when people seriously considered the Turing test passed reliably for the first time. But for the sake of this argument, it hardly matters what date years back we choose. There's been enough time since then to see the plateau.

> Most people in society connect AI directly to ChatGPT and hence OpenAI.

I'd double-check that assumption. Many people I've spoken to take a moment to remember that "AI" stands for artificial intelligence. Outside of tongue-in-cheek jokes, OpenAI has about 50% market share in LLMs, but you can't forget that Samsung makes AI washing machines, let alone all the purely fraudulent uses of the "AI" label.

> And there has been a lot of progress in image generation, video generation, ...

These are entirely different architectures from LLM/chat though. But you're right that OpenAI does that, too. When I said that they don't stray much from chat, I was thinking more about AlexNet and the broad applications of ML in general. But you're right, OpenAI also did/does diffusion, GANs, transformer vision.

This doesn't change my views much on chat being "not seeing the forest for the trees" though. In the big picture, I think there aren't many hockey sticks/exponentials left in LLMs to discover. That is not true about other AI/ML.

ChatGPT was not released to the general public until November 2022, and the mobile apps were not released until May 2023. For most of the world LLM's did not exist before those dates.

LLM AI hype started well before ChatGPT.

This site and many others were littered with OpenAI stories calling it the next Bell Labs or Xerox PARC and other such nonsense going back to 2016.

And GPT stories kicked into high gear all over the web and TV in 2019 in the lead-up to GPT-2 when OpenAI was telling the world it was too dangerous to release.

Certainly by 2021 and early 2022, LLM AI was being reported on all over the place.

>For most of the world LLM's did not exist before those dates.

Just because people don't use something doesn't mean they don't know about it. Plenty of people were hearing about the existential threat of (LLM) AI long before ChatGPT. Fox News and CNN had stories on GPT-2 years before ChatGPT was even a thing. Exposure doesn't get much more mainstream than that.

> LLM AI was being reported on all over the place.

No, it wasn't.

As a proxy, here's HN results prior to November, 2022 - 13 results.

https://hn.algolia.com/?dateEnd=1667260800&dateRange=custom&...

Here's Google Trends, showing a clear uptick May 2023, and basically no search volume before (the small increase Feb. 2023 probably Meta's Llama).

https://trends.google.com/trends/explore?date=today%205-y&ge...

https://trends.google.com/trends/explore?date=today%205-y&ge...

As another proxy, compare Nvidia revenues - $26.91bln in 2022, $26.97bln in 2023, $60bln 2024, $130bln 2025. I think it's clear the hype didn't start until 2023.

You're welcome to point out articles and stores before this time period "hyping" LLM's, but what I remember is that before ChatGPT there was very little conversation around LLM's.

If you're in this space and follow it closely, it can be difficult to notice the scale. It just feels like the hype was always big. 15 years ago it was all big data and sentiment analysis and NLP, machine translation buzz. In 2016 Google Translate switched to neural nets (LSTM) which was relatively big news. The king+woman-man=queen stuff with word2vec. Transformer in 2017. BERT and ELMo. GPT2 was a meme in techie culture, there was even a joke subreddit where GPT2 models were posting comments. GPT3 was also big news in the techie circles. But it was only after ChatGPT that the average person on the street would know about it.

Image generation was also a continuous slope of hype all the way from the original GAN, then thispersondoesnotexist, the sketch-to-photo toys by Nvidia and others, the avocado sofa of DallE. Then DallE2, etc.

The hype can continue to grow beyond our limit of perception. For people who follow such news their hype sensor can be maxed out earlier, and they don't see how ridiculously broadly it has spread in society now, because they didn't notice how niche it was before, even though it seemed to be "everywhere".

There's a canyon of a difference between excitement and buzz vs. hype. There was buzz in 2022, there was hype in 2023. No one was spending billions in this space until a public demarcation point that, not coincidentally, happened right after ChatGPT.

>In the big picture, I think there aren't many hockey sticks/exponentials left in LLMs to discover. That is not true about other AI/ML.

We do appear to be hitting a cap on the current generation of auto-regressive LLMs, but this isn't a surprise to anyone on the frontier. The leaked conversations between Ilya, Sam and Elon from the early OpenAI days acknowledge they didn't have a clue as to architecture, only that scale was the key to making experiments even possible. No one expected this generation of LLMs to make it nearly this far. There's a general feeling of "quiet before the storm" in the industry, in anticipation of an architecture/training breakthrough, with a focus on more agentic, RL-centric training methods. But it's going to take a while for anyone to prove out an architecture sufficiently, train it at scale to be competitive with SOTA LLMs and perform enough post training, validation and red-teamint to be comfortable releasing to the public.

Current LLMs are years and hundreds of millions of dollars of training in. That's a very high bar for a new architecture, even if it significantly improves on LLMs.

Seems like an arbitrary distinction.

I'd say Chain-of-Thought has massively improved LLM output. Is that "incremental"? Why is that more incremental than the move from GPT-2 to GPT-3? Sure you can say that this is when LLMs first passed some sort of Turing test, but fundamentally there was no technological difference from GPT-3 to GPT-4. In fact I would say the quality of GPT-4 unlocked thousands (millions?) more use-cases that were not very viable with the quality delivered by GPT-3. I don't see any reason for more use-cases to keep being unlocked by LLM improvements.

You saying —- with a straight face —- that post 2020 LLM AIs have made only incremental progress?

Yes. But they have also improved a lot. Incremental just means that the function is going up without breaking points. We haven't seen anything revolutionary, just evolutionary in the last 3 years. But the models do provide 2 or 3 times more value. So their pace of advancement is not slow.

The better you know a field the more it looks incremental. In other words, incrementalness is more a function of how much attention you pay or how deep you research it. Relativity and quantum mechanics were also incremental. Copernicus and Kepler were incremental. Deep learning itself was incremental. Based on almost identical networks from the 90s (CNN), which were using methods from the 80s (backprop) on architectures from the 70s (neocognitron) using activation functions from the 60s and the basic neuron model from the 40s (McCullough and Pitts), which was just a mathematization of observations in biology via microscopy integration with mathematical logic and electrical logic gates developed around the same time (Shannon), so it's just logic as formalized by Gödel and others and it goes back to Hilbert's program, which can be extrapolated from Leibniz etc. etc. It's not hard to say that "it's really just previous thing X plus previous thing Y, nothing new under the sun" to literally anything.

"It just suddenly appeared out of nowhere" is just a perception based on missing info. Many average people think ChatGPT was a sudden innovation specifically by OpenAI seemingly out of nowhere. Because they didn't follow it.

This is a sufficiently advanced science is indistinguishable from magic phenomenon.

The more you know about it, the less groundbreaking it is.

Yep, compared to beating the Turing test, the progress has been linear with exponentially growing investment. That's diminishing marginal returns.

Reply on news.ycombinator.com