Generative AI was a sort of paradigm shift, and can be developed into interesting tools that boost human productivity. But those things take time, sometimes decades to reach maturity.
That is not good for the get rich quick machine of Venture Capital and Hustle Culture, where quick exits require a bunch of bag holders.
You gotta have suckers, and for that Gen AI cannot be an "interesting technology" with good potential. It needs to be "the future", that will disrupt everything and everyone.
I disagree with 3.0, but perhaps that feels true for 4.0 or even 3.5 for some queries.
The reason is that when LLMs are asked questions whose answers can be interpolated or retrieved from their training data, they will likely use widely accepted human knowledge or patterns to compose their responses. (This is a simplification of how LLMs work, just to illustrate the key point here.) This knowledge has been refined and has evolved through decades of human experiments and experiences.
Domain experts of varying intelligence will likely come up with similar replies on these largely routine questions as well.
The difference shows up when you pose a query that demands deep reasoning or requires expertise in multiple fields. Then, frontier reasoning models like o3 can sometimes form creative solutions that are not textbook answers.
I strongly suspect that Reinforcement Learning with feedback from high-quality simulations or real environments will be key for these models' capabilities to surpass those of human experts.
Superhuman milestones, equivalent to those achieved by AlphaGo and AlphaZero between 2016 and 2018, might be reached in many fields over the coming years. This will likely happen first in fields with rapid feedback loops and highly accurate simulators.
I think that's the key threshold all these companies have been running up against, and crossing it would be the paradigm shift we keep hearing about. But they've been trying for years, and haven't done it yet, and seem to be plateauing
And then in OpenAI's case specifically- this tech has become commoditized really quickly. They have several direct competitors, including a few that are open, and their only real moat is their brand and Sam's fundraising ability. Their UX is one of the best right now, but that isn't really a moat
What? MSFT is worth 3.25 trillion without scaling beyond any users. All these AI companies need is for everyone to pay them $100 a year for a subscription and they have more than justified their valuation. Same strategy Microsoft uses
We had big progress in AI in last 2 years but have to take into account more than text token generation. We have image generation that is not only super realistic but you just text what you want to modify without learning complicated tools like ComfyUI.
We have text to speech and audio to audio that is not only very realistic and fluent with many languages but also can express emotions in speech.
We have video generation that is really more realistic every month and taking less computation.
There is big progress in 3d models generation. Speech to text is still getting improved and fast enough to run on phones reducing latency. Next frontier is how AI is applied for robotics. No to mention areas not sexy to end users but in application in healthcare.
Heck, I wrote myself my own personal radio moderator in a few hundred lines of shell, later rewritten in Python. As a simple MPD client. Watch out for a queued track which has albumart, and pass the track metadata + picture to the LLM. Send the result through a pretty natural sounding TTS, and queue the resulting sound file before the next track. Suddenly, I had a radio moderator that would narrate album art for me. It gave me a glimpse into a world that wouldn't have been possible before. And while the LLM is basically writing the script, the real magic comes from multimodal and great sounding TTS.
Much potential for really cool looking/sounding PoCs. However, what makes me worry is that there is not much progress on (to me) obvious shortcomings. For instance, OpenAI TTS really can't speak any numbers correctly. Digits maybe, but once you hand it something like "2025" the chance is high it will have pronounciation problems. In the first months, this felt like bad but temporarily acceptable. A year later, it feels like hilariously sad that nothing has been done to address such a simple yet important issue. You know that something bad is going on when you start to consider expanding numbers to written-out form before passing the message to the TTS. My girlfriend keeps joking that since LLMs, we now have computers that totally can not compute correctly. And she has a point. Sure, hand the LLM a tool to do calculations, and the situation improves somewhat. But it seems to be underlying, as shown by the problems of TTS.
Vision models have so many applications for me... However, some of them turn out to be actually unusable in practice. That becomes clear when you use a vision model to read the values off a blood pressure sensor. Take three photos, and you get three slightly different values. Not obviously made up stuff, but numbers that could be. 145/90, 147/93, 142/97. Well, the range might be clear, but actually, you can never be sure. Great for scene and art descriptions, since hallucinations almost fall through the cracks. But I would never use it to read any kind of data, neither OCR'd text nor, gasp, numbers! You can never know if you have been lied to. But still, some of the byproducts of LLMs feel like a real revolution. The moment you realize why whisper is named like that. When you test it on your laptop, and realize that it just transcribed the YouTube video you were rather silently running in the background. Some of this stuff feels like a big jump.
Human level object recognition can easily be trained up for custom use cases. Image segmentation is amazing. I can take a photo of a document and it’s accurately OCRd. 10-15 years ago that would be unfathomable.
I think current LLMs would give AI a much better reputation if they focused on non generative applications. Sentiment analysis, translation, named entity extraction, etc. these were all problems that data folks have been wrestling with that could very well be seen as “solved” and a big win for AI that businesses would b able onconfidently integrate into their workflows, but instead they went with the generative route and we have to deal with hallucinations and slop
However, while OCR done by vision models feels neat, I personally dont feel like it changed anything for me. I have been using KNFB Reader and later Seeing AI, and both have sufficiently solved the "OCR a document you just photographed" use case for me. They even aid the picture taking process by letting me know that a particular edge of the document is not visible.
Besides, I still don't understand the actual potential for hallucinations when doing OCR through vision models fully. I have a feeling there are a number of corner cases which will lead to hallucinations. The tendency to fill in things that might fit but aren't there is rather concerning. Talking about spelling errors and numerical data.
People are expecting it to get exponentially better, but these kind of innovations are more a inversed power law.
We look at CPUs or the transmission of digital data and these seem to have improved exponentially but these are rather exceptions and are composed of multiple technologies at different stages. Like how we went from internet through phone lines, to dedicated copper lines for data, to optic fiber straight into to people's homes.
Eg: look how the efficiency of solar cells has progressed over the last 50 years
The internet was adding a trillion dollars to the global economy by 2008 and the end of that rapid expansion, where AI is still sucking hundreds of billions a year into a black hole with no killer use cases that could possibly pay off its investment, much less begin adding trillions to the global economy.
And a decade before the web and internet explosion, PCs were similar, with a massive build out and immediate massive returns.
This excuse making for AI is getting old. It needs to put up or shut up because so far it's a joke compared to real advances like the PC and the Internet, all while being hyped by VC collecting companies as the arrival of a literal God.
This is not entirely true, or at least the trend is not necessarily less hallucination. See section 3.3 in the OpenAI o3 and o4-mini System Card[1], which shows that o3 and o4-mini both hallucinate more than o1. See also [2] for more data on hallucinations.
I think you're misremembering how 3.0 worked. Granted, the slope from 2.0 to 3.0 was very steep, but a ton of progress has happened in the past few years.
Also, insofar that they're researching a topic, they're not sufficiently critical. They are highly influenced by what they read, and they tend to take the results they're given more or less at face value.
For instance, I asked about a medical product for my girlfriend. I explicitly asked for a critical look at its efficacy, any studies in the area, etc.. and it seemed like 90% of the sources it considered were from the product's own website. It basically gave me an uncritical reporting of what they said about their own product.
Ie. better datasets.
Just recently I took a screenshot from a jira Burndown chart to write a description of the sprint progress for our stakeholders. Did it in one shot from a screenshot and got it right.
I wonder what “hugely more intelligent” would look like to you?
Also, in the (rather short) history of computing, what “hugely more X” has happened over the course of a couple of years?
They most certainly cannot
For context:
- GPT 2 was released in Feb 2019
- GPT 3 came out roughly 18 months later in 2020. It was a huge jump, but still not "usable" for many things.
- InstructGPT came out roughly 18 months later in early 2022, and was a huge advancement. This is RLHF's big moment.
- About 10 months later, ChatGPT is released at the end of 2022 as a "sibling" to InstructGPT. It's an "open research preview" at this point. This is around the time OpenAI starts referring to certain models as being in the "3.5 family"
- GPT-4 comes out in March 2023, so barely 2 years ago now. Huge jumps in performance, context window size, and it supports images. This is around the time ChatGPT hits 100 million users and is really becoming a reliable, widely adopted tool. This is also the same time that tools like Cursor are hitting the market, though they haven't exploded yet. Models are just now getting "good enough" for these kinds of applications
- GPT-4-Turbo comes out in November 2023, with way larger context windows and lower pricing.
- About 12 months ago, GPT-4o released, showing slightly increased performance on existing benchmarks over 4, but now with state-of-the-art audio capability support for something like 50 languages.
- 5 months ago, o1 releases. This is a big moment for scaling compute at test time, which is a major current research direction in ML. Shows huge improvements (something like 8x over 4) on some math/reasoning benchmarks. Within months, we have o3 and o4, which substantially improve these scores even further.
- In February of this year, we get 4.5, and then months later, the confusingly named 4.1, which shows improvements over 4o.
So to be clear, in 2019 we had an interesting research project that only a few people could tinker with.
18 months later, we had a better model that you could play with via an API, but was still a toy.
It takes more than two years to go from that to ChatGPT, and a few more months (nearly 3 years total) to get to the "useful" version of ChatGPT that really sets the world on fire. It took roughly 4 + 1/2 years to go from "novelty text generation" to "useful text generation".
In the 2 years since then, we've gotten multimodal models, a new class of reasoning models, baseline improvement across performance, and more. If anything, there is more fundamental research and wider variety of directions now (the kind of stuff that shifts paradigms) than before.
Yeah, sorry, but your view is not supported by data. There are billions of $ spent on tokens every year (between oAI, anthropic, goog, etc). This is not a toy anymore. People are using it, one way or another so it's useful to them - to the tune of billions of dollars. How useful it'll turn out to be in the abstract is still up for debate, but it is useful today.
If the product were to disappear tomorrow, I doubt the real economy would notice the loss of 0.05% in productivity.
[1] https://www.technologyreview.com/2025/02/25/1111207/a-nobel-...
That's not a viable business - that's still a VC funded growth mode startup.
Among knowledge workers in general, ChatGPT is used widely for basically any task that requires writing or researching.
This is obviously not going to be true for every single knowledge worker in every single role, and it seems that you don't find it particularly useful, but the volume of paying users is hard to dismiss out of hand.
It’s also improved my ability to analyze data, generating graphs and insights. But I still need to run that last mile myself, because I can’t fully trust its output. Same for web search actually, when I have a need to be comprehensive.
That's the step that still seems to be missing
At the same time, please don't react by breaking the site guidelines yourself or by putting down the community. It doesn't help; it only makes things worse.
If you know more than others do, or have something more constructive to say, that's great! Please share some of that so the rest of us can learn and converse. If you don't want to or don't have time, that's also fine, but please don't add putdowns to the thread.
It's always an option to downvote and/or flag comments that are breaking the rules, but if another commenter is wrong and you don't want to post a constructive correction, it's best to just move on.
It doesn't really matter how speculative the AGI bet is, their consumer AI business by itself is basically guaranteed to drown them in money. The only reason they're making losses at the moment is because they're choosing not to monetize their free tier users with ads, presumably since they don't need to make a profit and can prioritize growth.
But the moment they flip the advertising switch, their traffic will be both highly monetizable and ludicrously high margin.
But it seems pretty obvious that the math must be based on some incorrect assumptions. The unit inference costs of high quality LLMs are much lower than the unit costs of serving high quality web search queries. The LLM costs have also been decreasing rapidly -- 1000x in two years seems like a fair estimate -- while search engine costs haven't. (If anything they've been going up.)
And web search is obviously a business that is very profitable with the ad model, despite those higher unit costs.
> OAI will need more than ChatGPT, pro and "free" to be anything close to web search for ad revenue.
I don't understand this part at all. Revenue is proportional to traffic and ad rates. Why would the ad rates be lower for chatbots? Or why would their traffic obviously be lower?
This is also moving the goalposts. Search ad revenue is what, a $250B/year market. OpenAI doesn't anywhere near that much revenue to be fabulously profitable. A tenth of it would already be more than enough.
I think most people will continue to use Google and its Gemini-generated summaries at the top.
They already have half a billion MAUs. I'm pretty sure what I said is true already on that user base. It's not contingent on chatbots replacing search as a whole, all that's needed is their existing traffic having a decent proportion of monetizable queries (this seems basically certain) and for them to come up with ad formats that are effective without driving users away.
1. People won’t ultimate go download ChatGPT.app or use a website — they’re going to be using the functionality through structured services in iOS and Android, and it’s necessarily going to be under control of the OS vendors for security/privacy reasons. This doesn’t mean Apple and Google own the LLMs — there will be consumer choice for antitrust reasons if nothing else — but the operating system has to be a conduit for access to your data, and also for unified user experience. Which means advertising will be limited.
2. Say it does go the way you think it will — what prevents a real non-profit, open source LLM from taking it away from the commercial players? There really is no moat (other than money, energy, data center space).
- Go down to your local Ray-Ban store.
- Ask to play with a pear of the Meta glasses.
- Activate "Live AI mode"
- Have a real time video conversation with an AI which can see what you see, translate between languages, read text, recognize objects, and interact with the real world.
Contrary to your (potentially misremembered?) history, nothing at all like this was possible in 2019. I remember finetuning an early GPT-2 (before they even released the 2B model!) on a large corpus of Star Wars novels and being impressed that it would mention "Luke" when I ran the produced model! Now I wear it on my head and read restaurant menus with it. Use it to find my Uber (what kind of car is that?) Today I am building my raised garden beds out back and reading the various soil amendments I purchased, talking about how much bloodmeal to put over the hugelkultur layer, having it do math, and generally having a pair of eyeballs. I'm blind. The amount of utility I get out of these things is ... very hard to overstate.
If this's "moribund," sign me up for more decay.
Maybe it's me having an extremely low imagination, but that stuff existed for a while in the shape of google lens and the various vision flavor of LLMs, and I must have used them.... 3 times in years, and not once did I think "Gosh I wish I could just ask a question aloud while walking in the street about this building and wait for the answer". It's either important enough that I want to see the wikipedia page straight from google maps and read the whole lot or not.
> an AI which can read text, recognize objects, and interact with the real world.
I can already do that pretty well with my eyeballs, and I don't need to worry about hallucinations, privacy, bad phone signal or my bad english accent. I get that is certainly an amazing tools for people with vision impairments, but that is not the market Meta/OpenAI are aiming for and forcefully trying to shove it into.
So yes, mayyybe if I am in a foreign country I could see a use but I usually want to get _away_ from technology on vacation. So I really don't see the point, but it seems that they believe I am the target audience?
I see. Perhaps your eyeballs missed the part where I said I'm blind?
The entire purpose of my comment was to push back against this idea that AI is stuck in 2022. It's weird and nonsensical and seems disingenuous, especially when I say "here are things I can do now that I couldn't do before" and the general response is "but I don't need to do those things!"
I think anyone saying AI has no use is being willfully ignorant, but like every hype cycle before it since mobile (the last big paradigm shift), IMO it's going to result in a few useful applications and not the paradigm shift promised.
>>> I get that is certainly an amazing tools for people with vision impairments, but that is not the market Meta/OpenAI are aiming for and forcefully trying to shove it into.
I think a charitable reading of this thread is simply: AI as a large technology leap is still developing a business case that can pay for all the hype. Not to mention it's operating cost.
been the impact of OpenAI and meta glasses / headsets on the blind community at large?
Based on your statements it seems that the real value of AI is increasing the participation rate of visually impaired people in the global workforce.
If Elon or Sam can convince governments and insurance companies to pay for AI-powered glasses as a healthcare necessity maybe there’s a pathway forward for AI and VC class after all.
…maybe that’s the real game plan for Marc Andreesen, Kanye, Elon and the others.
They’re not really Nazi’s just early adopters choosing the “innovative freedom” promised by Emperor Palpatine and the Sith over the slow march of the Senators of the Republic.
Sorry, I’ve been watching too much Andor.
After I finished writing the message I realized that you were blind but I was too enraged to do the right thing and scrape the message. When I came back to my senses the anti-procastination had kicked in and I couldn't edit it.
I am sorry losing my temper, and I am very happy that this tech is able to be useful to you, and I think that it is a fantastic use case.
Now I need to get off the internet, it just brings the worst in me.
Ironically, this typo is very likely a result of AI dictation making a mistake. There are a lot of common misspellings in English, like "their" and "there", but I've never seen a human confuse "pair" and "pear".
So yeah, there are cool demos you can do that you couldn't five years ago. But whether any of those cool demos actually translate into something useful in day-to-day life where the benefits outweigh the costs and risks is far from clear.
I somehow agree with the op, that I don’t think I’m much closer to hiring chatGPT for a real job in 2025 than I was in 2022, but also you that there has been meaningful progress. And in particular, products that are transformative for disabled people are usually big improvements to the status quo for abled people too (oxo good grips being the classic example—transformative for people with arthritis, and generally just better for everybody else)
Every time I’ve encountered an AI first-line support agent I still find myself looking for the quickest escalation path to a real human just like before.
See, I always start with conversations about things I already know about, and they bullshit me enough that I'm wary of Gell-Mann Amnesia when asking them about things I don't know about. They output a lot of things that seem plausible but the way they blend fact and fiction with no regard for the truth keeps me extremely distrusting of them.
That is to say: after your conversation, did you ask for citations and go read the primary sources? Because if you did not, the model likely mislead you about Postgres in subtle ways.
Like "can a ps2 game use the ps1 hardware" it gave a noncommital, hallucinated answer. Then when asked to list sources it "searched the Internet" where all the links were from searches like "reddit ps2" etc.
It’s the “boring” stuff that’s interesting: automating drudgery work, a better way to research, etc.
I’ve been predicting for years that glasses — whether AR or VR — are and will remain niche. I don’t think most people want them.
Which makes sense, really. Buying an expensive pair of glasses so that I can point at an object and say “what is that” is a cool parlour trick but I can imagine very few scenarios where I’ve wished I had that functionality.
Realtime translation… absolutely a feature I’d use. For a week a year, max. IMO the killer, everyday application just isn’t there yet. I’m still not sure what it would be.
Even post dotcom bust there was still this generally shared understanding that you probably shouldn't meet people you met on the Internet and you shouldn't get in strangers cars. Today we use the Internet for the purpose of summoning strangers to us for the purpose of getting in their cars, and it's completely banal to do so.
These shifts take time and in the early-adopter stage the applications look half-baked, or solutions looking for a problem. That's because most of them are. The few that aren't will accrete over time until there's a solid base.
So, lots of people were meeting online in 1995. 5 years later couples who met on the internet were all over the place. When I moved to Silicon Valley in 2000, "I met my wife online" was hardly even a conversation starter. (I did meet my wife at The State of Insanity web chat boards in 1995.)
Also, from 1998-2008, the world invested precisely what it's now invested in the last decade of LLM AIs and was already adding over a trillion dollars annually to the global economy. AI, with that same spending over about the same amount of time, is still an economic black hole and no one has any answers that would satisfy anyone other than VC about how it's all going to pay for itself.
The internet, and particularly email, IM and the web, were going strong in the 90s and by the mid-2000s were absolutely dominant. This re-writing of history to suggest that AI deserves more runway to burn more hundreds of billions on things that will have no staying power like the trillion dollars in internet build out between '98-'08.
These things don't take time. The whole broadband internet took about a decade and about as much money as AI has gotten in this latest bubble. At least when the dot com bubble burst, we had infrastructure with staying power and real societal value, unlike AI which will leave us with next to nothing when its bubble pops.
If this is actually making software easier to make where is the software?
Do LLMs just not work for mobile development? What about native? Why is the only thing I'm actually seeing produced increasingly byzantine tooling for AI itself?
You are referencing a later period where regular folks were piling onto the internet and giving each other advice. And maybe that advice was geared towards novices and/or children, coming from other novices, and steeped in a lot of urban legend fears of the unknown. Even the academic types started to embrace these ideas, lending credence to the idea that their magical garden was now turning into hunting grounds.
But, imagine yourself in the early 90s, mapping your way (on paper, of course) to a trailer home in Silicon Valley or a secluded dirt road in the Santa Cruz mountains. You've never been there before, and your goal is to meet people you've never only seen as text on a screen. You assume they are real people but, technically, it could be all sock puppets.
I did those things! It wasn't a horror movie. Instead it lead to light socializing and drinking.
Maybe. What I remember is that the niche parts of the Internet remained niche. Everybody seemed excited about email, kids were excited about instant messaging, etc, and fundamentally that never changed; people have continued to use the Internet for communication. And the dotcom boom was a massive, stupid frenzy fueled by investors who had no idea what they were doing or what they were buying, and just wanted to make sure they didn't miss the next Microsoft, but the bubble was grounded in the accurate prediction that the web would become an essential part of business and an enormous money maker.
Does all of this sound right? Please correct me if I'm wrong. I find myself in the unexpected, uncomfortable position of wildly contradicting and undermining what I thought I understood: that people are terrible at predicting the future of technology.
Broadly speaking, in the 80s and 90s, the outlines of the future of technology was obvious and unmistakable: computers would become faster and more connected; that they'd take over or supplement a widening range of tasks and parts of our lives.
But it also can be hard to remember the way you understood the world in the past, or the way things were, because your understanding of the present overwrites it. Remember when VR was going to be huge? Remember when 3d chat was going to be the next big thing? Remember when TV was going to be hugely important in the classroom? Remember when social media was going to make our lives better?
People are making predictions regarding AI (or "AI") -- that it will do Everything Everywhere Very Soon, that it will lead to a 10x increase in worker output -- that strike me as absolutely, obviously wrong, even risible.
By and large, I think you're still right. But by the late '90s and early 2000s, the internet wasn't "the future of technology". It was already the present of technology.
Anyone paying attention could already see by 1994 or 1995 that the Internet was a) a big deal, and b) growing massively. The key at that point was not figuring out that the technology was going to be huge; it was figuring out which companies using that technology would be huge—ie, predicting with more granularity the future of the technology that was the World Wide Web.
And investors failed at that. Hard. That's what caused the dot-com bust.
Thid I'd why it sucks that glass was shot down when that technology will soar in a few decades.
People need to be more forward looking and open to change imo. It affects society in very negative ways. Imagine if people had planned for social media when it was first becoming a thing. We'd have safety nets, and an actual fucking plan.
There's been a stepwise jump in the capabilities of AI that's changed "products" from mostly fun to actually useful
This certainly provides benefit to those with limited vision, which is great. But that is a very small segment of consumers. Besides those, how many other people do you know who are actually _using these glasses_ in the real world?
Google Glass came out 10 years ago.
How does this personal anecdote relate to the observation that (so-called "tech") "[v]enture capital is moribund except for AI. AI is moribund except OpenAI."
Is venture capital involved with the Meta Ray-Ban glasses that we are advised to try for free.
does it? An anecdote from yesterday: My wife was asked to do a lit review as an assignment for nursing school. Her professor sent her an example of papers on the topic, with a brief "relevance" summary for each. My wife asked me for help as she was frustrated she couldn't find any of the referenced papers online (she's not the most adept at technology and figured she was doing something wrong). I took one look at the email from her professor and could tell just by the formatting that it was LLM generated (which model, I don't know, but obviously a 2025 model). The professor didn't say anything about using an LLM, and my wife didn't suspect that might be the case.
My wife and I did some Google Scholar searches, and _every_ _single_ _one_ of the 5 papers cited did _not_ exist. In 2 of the cases, similar papers did exist, but with different authors or a different title that resembled the fake "citation". The three others did not exist in any form - there were other papers on the same subject, sure, but nothing closely resembling the "citations" either in terms of authorship or title.
The interesting thing, to me, is how speculative OpenAI's bet is.
IIRC it was 2019 when I tinkered with the versions of GPT 2.0 that had web interfaces, and they were interesting toys. Then I began using ChatGPT since its launch, which was around Dec 2022, and that was a profound paradigm shift. It showed real emergent behavior and it was capable of very interesting things.
2019 - 2022 was three years. No hype, no trillions of dollars invested, but tremendous progress.
Now, there has been progress in the part ~three years in synthetic benchmarks, but the feeling with ChatGPT 4.5 today is still the same as it was with GPT-3/GPT-4 in 2022. 4.5/o3 doesn't seem hugely more intelligent then 3.0 -- it hallucinates less, and it's capable of running web searches and doing directed research -- but it's no paradigm shift. If things keep progressing the way they're going, we'll get better interfaces and more tools, but it's far from clear that superintelligence (more-than-human insight, skill, and inventiveness,) is even possible with LLMs.