36 votes

OpenAI’s H1 2025: $4.3b in income, $13.5b in loss

64 comments

  1. [53]
    teaearlgraycold
    Link
    HN users are saying that once they start pumping ChatGPT full of ads this could be profitable. I'm not so sure but we will see. I don't really use LLMs much outside of coding. Honestly I don't...

    HN users are saying that once they start pumping ChatGPT full of ads this could be profitable. I'm not so sure but we will see. I don't really use LLMs much outside of coding. Honestly I don't think the ChatGPT-like products have much staying power. LLMs are better when tightly integrated into products like with Cursor.

    19 votes
    1. [15]
      Eji1700
      Link Parent
      The main market I see outside of coding (or perhaps tangential) is help documentation. Its a generic interface to your help documentation that you wrote forever and no one reads. "how do I do...

      The main market I see outside of coding (or perhaps tangential) is help documentation.

      Its a generic interface to your help documentation that you wrote forever and no one reads. "how do I do blah" typed into google or whatever gets people lost on garbage fast and doesn't iterate well when you say "but what about blah in my case?"

      LLM's do handle that decently well. Not great, but given that most companies are sitting on MILLIONS on optimizations if they could just get their userbase to use 10% more of their O365 tools or whatever, it might pencil out.

      The problem is, how do you market that? Are you going to make them pay for easier to use, one and done, help documentation? Do i not just have one account then for trainee's/occasional use? There's no fucking way it works if you start shoving ads in, and you STILL run the risk of "look it worked the last 10 times so I just did what it said again? How was I supposed to know it did the math wrong and now we've got a 40% shortfall?" thing.

      16 votes
      1. [9]
        Lyrl
        Link Parent
        I am really looking forward (hoping!) my work (large manufacturing plant) does this for our internal process documentation. It's really helpful if you find the right piece and that piece is up to...

        I am really looking forward (hoping!) my work (large manufacturing plant) does this for our internal process documentation. It's really helpful if you find the right piece and that piece is up to date, but both finding and updating are questionable, and I believe an LLM could help.

        8 votes
        1. [6]
          tauon
          Link Parent
          At this point, I’m fairly convinced that even after the stock/investment market “AI” bubble comes crashing down, one of the few applications that’ll definitely still survive just for the sheer...

          At this point, I’m fairly convinced that even after the stock/investment market “AI” bubble comes crashing down, one of the few applications that’ll definitely still survive just for the sheer value it provides to orgs is RAG, and potentially a few consulting firms specialized to implement the tech. (To be clear, in LLM-based applications – of course the medical types of cancer-from-xray detection AI, trained on imagery not words, will survive either way, its value is also too good to give up.)

          Almost every company has some form of internal docs or wiki, and (until now) almost every company’s internal search for that knowledge has been poor in the median.

          8 votes
          1. chocobean
            Link Parent
            I would become a far happier user if RAG LLM can highlight text from non-generated exact document section when asked. It has a quite good guess on where info might be, but I need to read it to...

            I would become a far happier user if RAG LLM can highlight text from non-generated exact document section when asked. It has a quite good guess on where info might be, but I need to read it to make sure.

            I would also be happier to live in a world where LLM responses are always phrased as "probably x" or "I'm guessing y" etc.

            11 votes
          2. [4]
            teaearlgraycold
            Link Parent
            If you can believe it, Google has pretty good internal search.

            If you can believe it, Google has pretty good internal search.

            5 votes
            1. [3]
              skybrian
              Link Parent
              It used to be pretty mediocre (other than code search) so I’m glad to hear they finally fixed it after I left.

              It used to be pretty mediocre (other than code search) so I’m glad to hear they finally fixed it after I left.

              2 votes
              1. chocobean
                Link Parent
                I hope it's better than their public search, which is becoming more unusable by the day :/

                I hope it's better than their public search, which is becoming more unusable by the day :/

                4 votes
              2. teaearlgraycold
                Link Parent
                Well I don’t know if it got up to your standards. But it’s good compared to most companies. Code search was much better.

                Well I don’t know if it got up to your standards. But it’s good compared to most companies. Code search was much better.

                2 votes
        2. [2]
          sparkle
          Link Parent
          My organization has internal documentation spread across so many areas. GitHub, confluence, Jira, backstage, inherited wikis from acquisitions, Jim's "don't delete, impt docs" folder, etc. I'm...

          My organization has internal documentation spread across so many areas. GitHub, confluence, Jira, backstage, inherited wikis from acquisitions, Jim's "don't delete, impt docs" folder, etc. I'm half tempted to join our hackathon this year just to pitch a documentation oriented LLM to help organize all this. Sadly the hackathon will mostly be about "how can we use AI to gouge benefit customers". I guess this still falls under that umbrella lol, just not in the way management wants

          not just organization but, like you said, keeping it up to date as well! GitHub is actually pretty good at automatically updating documentation when setup right, but the rest are just generally unmaintained and unhelpful.

          Nobody wants to do the work to update and maintain it because we're all overworked (partially because of poor documentation) so an LLM would be an excellent tool in this situation

          5 votes
          1. Minori
            Link Parent
            If all the tools have a search interface, wiring them into an LLM as tools for MCP usage can be extremely useful. I'd still prefer a good search engine, but we take what we can get.

            If all the tools have a search interface, wiring them into an LLM as tools for MCP usage can be extremely useful. I'd still prefer a good search engine, but we take what we can get.

      2. [3]
        first-must-burn
        Link Parent
        One of the things I thought about having this area: I use an LLM to write documentation after I finish my code. I say "update the documentation to reflect these code changes" or more usually...

        One of the things I thought about having this area:

        I use an LLM to write documentation after I finish my code. I say "update the documentation to reflect these code changes" or more usually "generate the documentation" because it was completely missing. I still review it, but the usual problem with it is that it's accurate but it is just too verbose.

        Even before llms, I wrote good but very long documentation. And my experience with users is generally that people won't read it. If you say "the answer's in the doc" and the doc is 10 pages long, they just say "show me where" or "show me how". Which again is a perfect use case for an llm.

        So I think, why not cut out the middle step entirely. I think that an llm that had the code base as rag could probably answer questions pretty well from a user point of view.

        And certainly for developers, it could do a great job of answering questions about the API. I can almost imagine rag/mcp as a service, where you subscribe to the documentation service for the libraries you specifically use in order to make your code generation LLM very smart about those apia specifically. Making up an API call that doesn't exist is one of the most common hallucinations I see.

        In both cases, you'd be relieved from the pain of maintaining documentation across multiple versions because the llm can know what version the person asking has and generate answers specific to that version.

        8 votes
        1. skybrian
          Link Parent
          I expect that LLM’s will be the primary readers of documentation going forward, which is why some people recommend putting it all into one big Markdown file and not getting fancy. (Other than API...

          I expect that LLM’s will be the primary readers of documentation going forward, which is why some people recommend putting it all into one big Markdown file and not getting fancy. (Other than API documentation that’s best put near the code.)

          6 votes
        2. Minori
          Link Parent
          In my experience, the LLM generated docs are subtly wrong in ways that are really annoying to discover. I don't feel great about feeding LLM docs into another LLM to generate increasingly...

          In my experience, the LLM generated docs are subtly wrong in ways that are really annoying to discover. I don't feel great about feeding LLM docs into another LLM to generate increasingly inaccurate summaries...

          2 votes
      3. creesch
        Link Parent
        Generally speaking though, documentation itself isn't enough to train LLMs on. At least not to the degree that they can semi reliably answer "how do I do X". Overly simplified they "learn" best...

        Generally speaking though, documentation itself isn't enough to train LLMs on. At least not to the degree that they can semi reliably answer "how do I do X".
        Overly simplified they "learn" best from a wide array of examples and explanations.

        This is actually reflected in how well they currently are able to handle certain languages and technologies. The more popular they are the better models are with them. They are more likely to be able to output something related to Javascript compared to something that has less training material like C. Similarly they will be more likely to give a semi decent React answer compared to frameworks with a much smaller user base.

        It's not to say they can't do it, RAG based systems can do fairly well. But the quality will not be the same. Which means that the risk of hallucinations, which is already there, will be bigger.

        It also requires the documentation to be entire up to date and correct. Where with a wider training base gaps will be filled by examples.

        8 votes
      4. teaearlgraycold
        Link Parent
        If it's just there for semantic search then the onus is on the documentation, not the AI. I really like semantic search as a feature and wish it was more prevalent.

        If it's just there for semantic search then the onus is on the documentation, not the AI. I really like semantic search as a feature and wish it was more prevalent.

        4 votes
    2. [11]
      raze2012
      Link Parent
      Ads for a black box oracle sounds like the exact dystopia Silicon Valley would laugh at 10 years ago. Didn't take much to sell their souls, I guess. Reality is stranger than fiction, though. Seems...

      Ads for a black box oracle sounds like the exact dystopia Silicon Valley would laugh at 10 years ago. Didn't take much to sell their souls, I guess.

      Reality is stranger than fiction, though. Seems instead that ChatGPT is trying to turn into Tiktok to get that money back: https://andrewmayne.com/2025/10/04/sora-the-chatgpt-moment-for-generative-video/

      Having your entire social media become AI Slop feels like yet another omen dozens of sci-fi stories warned against.

      15 votes
      1. st3ph3n
        Link Parent
        I still use facebook for some things (I know, but it is unavoidable in some cases) and the amount of sensational AI slop videos it keeps trying to push into my face is disturbing.

        I still use facebook for some things (I know, but it is unavoidable in some cases) and the amount of sensational AI slop videos it keeps trying to push into my face is disturbing.

        4 votes
      2. [9]
        skybrian
        Link Parent
        Hacker News readers will say stuff but that doesn’t mean it’s true. OpenAI isn’t doing ads as such, not yet anyway. It’s closer to doing product recommendations and buy links. They want to make...

        Hacker News readers will say stuff but that doesn’t mean it’s true. OpenAI isn’t doing ads as such, not yet anyway. It’s closer to doing product recommendations and buy links. They want to make shopping using ChatGPT a thing it can do when people request it.

        2 votes
        1. [7]
          ep1032
          (edited )
          Link Parent
          I honestly think that AI is going to make the concept of an advertisement outdated. Like, sure, advertisements will still exist, because why not. But right now, we are living in a world where it...
          • Exemplary

          I honestly think that AI is going to make the concept of an advertisement outdated. Like, sure, advertisements will still exist, because why not.

          But right now, we are living in a world where it is technologically too difficult to do subtle product placement. Product placement sits like a sore thumb in most places where it is tried.

          Similarly, it is extremely difficult to change, dynamically, the way people communicate. If you want to make a movie script more right or left leaning, or softer on big oil or whatever, you need to pay people to rewrite the script.

          And the people producing that product are going to push back on such requests. Partly because each of these requests take time to implement, but more fundamentally because right now we differentiate between advertising and actual products. We do this, because producing a product requires capital, and advertising undermines trust in that product. So there is a built in social force for people to protect the investment they have made in their product and their brand's credibility by ensuring the consumer can differentiate between the product and advertising related to that product.

          But AI undermines all three of these things.

          And it introduces the ability to manipulate all three of these things on a much higher abstraction level.

          For example, right now if CoCa-Cola wants a product placement in a movie, then they could ask a character to buy a coke in a scene where they go to a bar, and characters could potentially do so. This requires some coordination during the filming to set up, and as a result a movie could only realistically do so many requests like this without bogging down production. But in a post-AI era, how much do you think that Coke would be willing to pay, to update a movie pre-release so that every drink in every background shot is carbonated? How much would they pay to ensure that such a filter was run on every movie that comes out this year?

          How much would Fox pay to change every TV in the background of every shot to have their color scheme, even if too blurry to see in shot?
          Do you think Ford would be willing to pay to increase the percentage of automobiles on every show shown on ABC to be 50% pickup trucks? Would Toyota pay for it to be 50% sedans?
          What if you could automatically update every time a radio host or newspaper editor or youtube influencer used a metaphor about cars, to instead automatically say "Ford F-150"?
          Do you think a political campaign would be interested in slowing down text messages in key counties only for people who have discussed more liberal or conservative topics with friends in the last 6 months leading into an election? Facebook and twitter are already hiding or promoting user posts based on this metric, how different would it be to do it to interpersonal communication?
          If advertisements themselves can be AI generated right before being shown to an individual person, how much could a marketing firm get a hotel company to pay to ensure that whenever a person is shown in an ad, that they are shown standing in a hilton, and not a hayatt?
          We can already pay Google to prioritize specific search results, how much would changing the language it uses in its responses be worth?
          What if you can change the language, for different socioeconomic groups?
          What if you can change the language those socioeconomic groups use themselves? Hell, just start with autocorrect. Would Ford be interested in ensuring that 20% of the time, when you type car for the next 3 months, your phone autocorrects to the word pickup?

          Net neutrality was never passed in this country, so what if when you sent a message, your message was just updated automatically after you sent it, and you never knew? The person you sent the message to wouldn't find it weird that you said pickup instead of car, how would you ever know? Surely, doing that would be valuable, wouldn't it? What if we only did it 5% of the time? Would anyone notice? Would that change spending habits? Would that flip elections?

          What could be a more effective form of advertising than changing the language people use, the language people hear, and the way the world they see around them looks?

          3 votes
          1. [2]
            mat
            Link Parent
            Um. Sorry to break it to you but product placement has been mutable digitally for years now. For example, different regional releases of films and TV have different local products matted into the...

            it is technologically too difficult to do subtle product placement. Product placement sits like a sore thumb in most places where it is tried.

            Um. Sorry to break it to you but product placement has been mutable digitally for years now. For example, different regional releases of films and TV have different local products matted into the back and foregrounds all the time. It's common practice.

            You only think the subtle stuff doesn't exist because you don't notice it. You're not supposed to notice it, that's not an accident!

            9 votes
            1. Lia
              Link Parent
              Sure, but the recent developments in generative AI enable it to go to a completely different level and scale. As an analogy, the fact that people used to read smutty magazines doesn't magically...

              product placement has been mutable digitally for years now.

              Sure, but the recent developments in generative AI enable it to go to a completely different level and scale.

              As an analogy, the fact that people used to read smutty magazines doesn't magically make all online porn harmless in every conceivable scale, scope and style.

          2. [3]
            skybrian
            Link Parent
            One way advertising works is as a reminder to the already presuaded. If someone who already likes McDonalds is reminded of them, maybe they'll decide to eat there for dinner? For these people, an...

            One way advertising works is as a reminder to the already presuaded. If someone who already likes McDonalds is reminded of them, maybe they'll decide to eat there for dinner? For these people, an appropriately timed reminder could possibly work. And sure, subtle reminders could work, but I think advertisers are generally more blatant about it, for example, by sending notification spam.

            For AI Chat, though, bringing up irrelevant stuff when you ask a question is pretty blatant and I see much more promise in AI becoming a capable shopping assistant, for when you are researching what to buy to fulfil a need.

            3 votes
            1. [2]
              Lia
              Link Parent
              Another way, which has been illegal for decades due to being so effective, is subliminal marketing. IIRC, methods included flashing a picture of some product to unsuspecting audiences while they...

              One way advertising works is as a reminder to the already persuaded.

              Another way, which has been illegal for decades due to being so effective, is subliminal marketing. IIRC, methods included flashing a picture of some product to unsuspecting audiences while they were consuming other content, so quickly that they couldn't detect having been shown anything at all.

              I wouldn't be surprised if the most profit-hungry businesses were constantly trying to invent equally effective ways that aren't banned yet, a task that gen AI seems ideally suited to, and that goes for both inventing and implementing.

              For AI Chat, though, bringing up irrelevant stuff when you ask a question is pretty blatant

              Why would they do it in a way that seems blatant when they can find ways for it to go undetected?

              I've received unsolicited advice to stop using TimeMachine, because it "just sucks", and invest in a third party solution instead when I was asking for advice on TM configuration. My problem was easily solved and absolutely didn't require switching to another system. The reason I was even asking was because I'm not tech savvy at all. If I hadn't received education in media literacy, I could have easily just taken that advice at face value due to my lack of tech know-how. I might have even felt like the LLM helped me and saved me tons of frustration when it got me out of using a sucky system.

              Manipulating different users in different ways, highly tailored to each user's individual personality and weaknesses, hasn't been technically feasible in the past but now that it is, it's definitely something aggressive advertisers are interested in.

              As a tangential note: Netflix is now optimising some of their key content for something they call "second screenability" - meaning when you're on your first screen (phone, tablet), you should be able to follow the story your tv is showing in the background. At first I didn't understand why they would want more viewers who aren't even really watching - that sounds like an awfully unappealing consumer group to present to advertisers. However, seeing as advertising is a lot more effective, at least in come cases, when the target audience is unaware, I'm starting to suspect that this is the main driver behind this development. After all, product placement etc. can't be banned as subliminal marketing if it's simply part of a tv show, as is customary these days. But if people aren't actually paying attention to that show while still frequently exposing themselves to it, then the user may turn a benign ad into a subliminal one by their own behaviour. It would be very hard to legislate against that.

              1. skybrian
                Link Parent
                Here’s a BBC article about subliminal advertising from ten years ago: Does subliminal advertising actually work? I’m not too impressed with the study, but it didn’t find much effect: I don’t think...

                Here’s a BBC article about subliminal advertising from ten years ago:

                Does subliminal advertising actually work?

                Psychologists have long agreed that flashing words too quickly for the conscious mind to register can have some limited effects in the lab.

                And at the University of Utrecht in 2006, a team of experimental social psychologists, Johan Karremans, Jasper Claus and Wolfgang Stroebe, did manage to make subliminal advertising itself work - in strict laboratory conditions, provided a series of limiting factors are in place.

                Their work suggested that subliminal advertising was only effective with products that people knew of and somewhat liked. The flashes made the brand name more '"cognitively accessible", their theory went, so it wouldn't work with very high-profile brands - you couldn't make a brand like Coca-Cola much more familiar to people than it already is.
                They replicated their results, and published their findings in the Journal of Experimental Social Psychology.

                But the crucial question, raised by Vicary's dubious claims, and never finally settled, is this - can you take all this out of the lab, beyond its strict controls, and reproduce it in the messiness of real life, on a mass scale?

                No-one, apparently, has attempted this since the 1950s. So, as part of a BBC Radio 4 documentary, we decided to carry out a public test.

                I’m not too impressed with the study, but it didn’t find much effect:

                The participants, members of the audience of Radio 4's science show The Infinite Monkey Cage, were then offered a choice of two drinks - Lipton Iced Tea or a brand of mineral water - and asked to complete a questionnaire.

                Strikingly, there was no significant effect.

                For all participants, a few more people in the test group picked Lipton, but not enough to be statistically significant. When we removed those likely to have been immune to the subliminals - ie those who would have picked Lipton anyway, and those who dislike it and would never pick it, slightly more people in the control group picked Lipton, but this difference was not significant either.

                I don’t think this is the last word, but my assumption is the subliminal advertising is less effective than regular advertising. It might be worth a more in-depth investigation, though; this is just the first article I found doing a search.

                1 vote
          3. Minori
            Link Parent
            Citation needed. Most modern advertising is subtle mind tricks to make you associate products with good things or develop addictive habits. Depending on your definition of ad, most packaging is an...

            and advertising undermines trust in that product. So there is a built in social force for people to protect the investment they have made in their product and their brand's credibility by ensuring the consumer can differentiate between the product and advertising related to that product.

            Citation needed. Most modern advertising is subtle mind tricks to make you associate products with good things or develop addictive habits. Depending on your definition of ad, most packaging is an advertisement.

            Take Coke's campaigns that are designed to make people associate happy times with friends and family with Coca-Cola products.

            For subtle advertising, consider all of the product placements in major movies. Maybe you don't notice it, but the protagonist puts on a jacket from a specific brand before her coolest scene. The bartender only stocks one brand of beer or all the liquor comes from one conglomerate. It's subtle, yet it works.

        2. raze2012
          Link Parent
          Sora is invite only right now. But unless they announce some sort of premium subscriptions, it seems like ads are inevitable once they open it up.

          OpenAI isn’t doing ads as such, not yet anyway.

          Sora is invite only right now. But unless they announce some sort of premium subscriptions, it seems like ads are inevitable once they open it up.

          1 vote
    3. [4]
      nocut12
      Link Parent
      I think translation will stick around as an important use case. Quality improved drastically post-LLM, and I've definitely gotten some really impressive results. It's obviously far from perfect...

      I think translation will stick around as an important use case. Quality improved drastically post-LLM, and I've definitely gotten some really impressive results. It's obviously far from perfect and I think human-made localizations will remain the gold standard, but there's incredible amounts of stuff that will simply never be translated otherwise -- things like commentary for sports, many newspapers/magazines, niche content from small publishers, etc.

      I don't think this is really a huge opportunity for tech companies (at least not as huge as you'd need to make up for 13.5 billion...), but it feels like it could be a big deal for rightsholders. I think there's an honest chance this stuff could mean a kinda-sorta end of the language barrier in non-face-to-face scenarios in the not too distant future, which would be pretty crazy.

      9 votes
      1. [3]
        teaearlgraycold
        Link Parent
        Also face to face. Apple is starting to work on that.

        Also face to face. Apple is starting to work on that.

        1 vote
        1. Lyrl
          Link Parent
          I think Apple is working on optimizations with voice-to-voice translation, but on travel YouTube channels it's already common to see them talking with locals through an app (usually...

          I think Apple is working on optimizations with voice-to-voice translation, but on travel YouTube channels it's already common to see them talking with locals through an app (usually voice-to-text). My mom recently had a car key copied by a person who spoke very little English, and used a translation app. She was satisfied by the experience.

          3 votes
        2. Minori
          Link Parent
          Per usual, Apple are a bit behind the curve on that. Google Translate has supported live conversation translations for a few years. They're okay.

          Per usual, Apple are a bit behind the curve on that. Google Translate has supported live conversation translations for a few years. They're okay.

    4. [7]
      infpossibilityspace
      Link Parent
      I'd agree that targeted integration is probably more useful than a generic chatbot, but I'm curious if there's much difference between the difference chat models? Why wouldn't people switch to a...

      I'd agree that targeted integration is probably more useful than a generic chatbot, but I'm curious if there's much difference between the difference chat models? Why wouldn't people switch to a different model if one of them starts spamming ads? I don't use ai so genuinely have no idea what a given model's USP is and if that difference is big enough to make people put up with increased ads.

      8 votes
      1. teaearlgraycold
        Link Parent
        I imagine the only thing that would keep people around would be the chat app's knowledge of the user and its supplemental tools - like image generation. Generally the top models between Anthropic,...

        I imagine the only thing that would keep people around would be the chat app's knowledge of the user and its supplemental tools - like image generation. Generally the top models between Anthropic, OpenAI, and Google are close in performance. If ads end up being necessary then Google (possibly Meta) will absolutely be the winner. They can both subsidize their LLMs indefinitely thanks to their existing ad business and already have the data to make what ads they do inject higher quality, which would mean they need less of them.

        6 votes
      2. [4]
        EgoEimi
        Link Parent
        There probably is little difference when it comes to easy tasks like asking how-to's for household tasks or repairs, analyzing popular novels, etc. The differences reveal themselves on hard use...

        There probably is little difference when it comes to easy tasks like asking how-to's for household tasks or repairs, analyzing popular novels, etc.

        The differences reveal themselves on hard use cases, like math, engineering, and other abstract reasoning tasks.

        Terrence Tao reported that he found ChatGPT to be a "significant time saver" in engaging it in an extended conversation to help him answer a MathOverflow question. Moreover, he reports he encountered no hallucinations — but he thinks that's because he had a good idea of what he wanted and was very specific and detailed in his prompting.

        ChatGPT is a double-edged sword. In the hands of newbies, it's going to create slop. But it's a force multiplier for skilled operators who are already knowledgeable in their domains. Personally, it has at least 2x'd my productivity; I'm accomplishing more and better.

        I highly doubt they're going to put in ads. It's more likely they're going to create knowledge-work-specialized professional models (SWE-specialized GPT, med-specialized GPT, finance-specialized GPT, etc.) with ironclad data privacy agreements to sell to companies at enterprise ($$$) rates for employee use.

        Enterprise, not consumer, is always where the serious money is.

        2 votes
        1. [3]
          skybrian
          Link Parent
          There is serious money to made in the enterprise market, provided businesses are convinced it’s effective. From early results that’s unclear, but perhaps it’s a skill issue. Enterprise and...

          There is serious money to made in the enterprise market, provided businesses are convinced it’s effective. From early results that’s unclear, but perhaps it’s a skill issue.

          Enterprise and consumer aren’t independent, though. It seems like ChatGPT is more consumer-oriented than other AI companies (such as Anthropic), and they also have the most well-known brand. It probably helps for enterprise sales, too. People will want to use the AI chat they’re familiar with.

          1 vote
          1. [2]
            Minori
            Link Parent
            I'm unconvinced these are "early results" considering GPT4 was released a couple years ago. Professionals have had plenty of time to integrate LLMs.

            I'm unconvinced these are "early results" considering GPT4 was released a couple years ago. Professionals have had plenty of time to integrate LLMs.

            1. skybrian
              Link Parent
              Sure, many things have been tried, but sometimes it takes years of refinement for new kinds of software to evolve. For example, early web browsers and IDE’s were promising but pretty crap. I think...

              Sure, many things have been tried, but sometimes it takes years of refinement for new kinds of software to evolve. For example, early web browsers and IDE’s were promising but pretty crap.

              I think basic AI chat is pretty mature at this point, but there’s not yet a consensus on the best way to integrate LLM’s into software development tools. Companies are trying lots of different variations. Software developers have different way of using these tools, too, and there isn’t a consensus on best practices.

              For other industries, I don’t know, but maybe there are similar issues?

              1 vote
      3. slade
        Link Parent
        I think the only way to make LLMs into sustainable products is with highly specialized models built and maintained by human domain experts. Right now the low barrier to entry means tons of low...

        I think the only way to make LLMs into sustainable products is with highly specialized models built and maintained by human domain experts. Right now the low barrier to entry means tons of low effort slop, but over time I suspect we will see high quality products emerge based on specialty LLMs.

        1 vote
    5. json
      Link Parent
      I can see the advertising be a very insidious sponsorship bias in any results. A 'common person' use case of conversational search is to be the travel agent: help me plan out a holiday in Fiji....

      I can see the advertising be a very insidious sponsorship bias in any results.

      A 'common person' use case of conversational search is to be the travel agent: help me plan out a holiday in Fiji.
      With collated reviews of hotels that I could see obtained from, sorted by and filtered through the advertising engine.

      Comparing car models, cpus, phones, etc.

      Easier to sway opinion with a more product placement type of ad than having 'relevant ads' adjacent to the conversation.

      4 votes
    6. [6]
      CptBluebear
      Link Parent
      Well they're going to allow direct sales (from Etsy if I'm not mistaken?) which should boost their revenue if it proves to be successful. Very much a shake the tree and see what falls out...

      Well they're going to allow direct sales (from Etsy if I'm not mistaken?) which should boost their revenue if it proves to be successful. Very much a shake the tree and see what falls out situation though!

      3 votes
      1. [3]
        chocobean
        Link Parent
        I feel like we've already tried this with "Alexa buy toilet paper". Human beings have very little patience for errors they didn't create. Multiplied by how easily we get mad when we lose time and...

        I feel like we've already tried this with "Alexa buy toilet paper". Human beings have very little patience for errors they didn't create. Multiplied by how easily we get mad when we lose time and money. I don't see this having any possibility of success beyond the extremely limited case of "buy exactly this item I have previously ordered" which is just a single button option we already have.

        10 votes
        1. CptBluebear
          Link Parent
          I agree, I'm sceptical this iteration will be successful. I do think that direct sales from a hyper personalized system is the monetized future they're aiming for, and likely the only one that...

          I agree, I'm sceptical this iteration will be successful. I do think that direct sales from a hyper personalized system is the monetized future they're aiming for, and likely the only one that will have a shot at profitability with the state LLMs are in right now. Considering the most "advanced life changing" stuff they can think of right now is... The same short form videos but automated.. I'm not very optimistic.

          6 votes
        2. mat
          (edited )
          Link Parent
          I cannot count the number of times I've had a number of specific requirements for buying something and wished I had a smart machine agent capable of finding and sorting a list which matches that...

          I cannot count the number of times I've had a number of specific requirements for buying something and wished I had a smart machine agent capable of finding and sorting a list which matches that for me. I spend so much time digging around in specs and descriptions which are not arranged in a standard way to try and find the item which ticks all the boxes I want ticking.

          For example this week: "find me a pair of glasses which have white or clear frames, arms which are a little interesting, probably a square-ish shape but definitely not round, zero rake and are available from a dispensing opticians who can fit ultra-thin aspheric varifocal lenses"

          Of course just integrating a chatbot into Etsy won't solve that problem. But maybe later someone else will.

          2 votes
      2. [2]
        teaearlgraycold
        Link Parent
        I really don’t see that being a high value integration. But their planned integration with Shopify will at least be lucrative for partners.

        I really don’t see that being a high value integration. But their planned integration with Shopify will at least be lucrative for partners.

        3 votes
        1. CptBluebear
          Link Parent
          I find Etsy to be questionable too. Rather the precedent and how they may integrate more and more of those deals if the Etsy deal is successful. It's the logical next step, or even end game, to...

          I find Etsy to be questionable too. Rather the precedent and how they may integrate more and more of those deals if the Etsy deal is successful. It's the logical next step, or even end game, to direct sell to customers. If there's anything in AI that will turn a profit, because not a single one of them is profitable right now, it'll be direct sales.

          4 votes
    7. PuddleOfKittens
      Link Parent
      "Hey chat, what should I do now?" "You should drink a Pepsi." ChatGPT add have unparalleled potential, because lots of people treat AI like a person they trust.

      "Hey chat, what should I do now?"

      "You should drink a Pepsi."

      ChatGPT add have unparalleled potential, because lots of people treat AI like a person they trust.

      2 votes
    8. [7]
      ackables
      Link Parent
      ChatGPT gets way more useful the longer you use it. For example, 6 months ago I was working on a project and asked ChatGPT for some help understanding how to work with a new framework I was using....

      ChatGPT gets way more useful the longer you use it.

      For example, 6 months ago I was working on a project and asked ChatGPT for some help understanding how to work with a new framework I was using. I recently wanted to update my resume and tailor it to a specific job listing, so I uploaded my current resume to ChatGPT and asked it to reword my bullet points to reflect the details of the job listing. It actually added a new entry for the project I was working on 6 months ago even though I didn’t think to ask and highlighted the parts that were relevant for the job I was applying to.

      Having an LLM that knows details about your life is incredibly useful as a “second brain” that can remind you of past details that may be relevant to what you are trying to accomplish today.

      1 vote
      1. [3]
        TurtleCracker
        Link Parent
        I’d only be willing to do this if the memory (and ideally the LLM) was local. I have memory turned off.

        I’d only be willing to do this if the memory (and ideally the LLM) was local. I have memory turned off.

        7 votes
        1. [2]
          ackables
          Link Parent
          I’m less worried since I’m on an edu license that doesn’t allow OpenAI to use my chat history for anything, but I would prefer a local LLM in the future.

          I’m less worried since I’m on an edu license that doesn’t allow OpenAI to use my chat history for anything, but I would prefer a local LLM in the future.

          1 vote
          1. Lia
            Link Parent
            I have history turned off but it's still being recorded. How do I know? Because it leaked into my chat. In a new chat, I was asking about something pertaining to my laptop and it went: "You could...

            doesn’t allow OpenAI to use my chat history for anything

            I have history turned off but it's still being recorded. How do I know? Because it leaked into my chat. In a new chat, I was asking about something pertaining to my laptop and it went: "You could do this on your M2 Mac Mini instead, if you still have it". That should be impossible if OpenAI was actually doing what they say they're doing.

            The point being: if you have no way to verify they are actually adhering to their license terms, then it's probably good to keep in mind the terms may be disregarded at any time, with or without your knowledge.

            1 vote
      2. [3]
        skybrian
        Link Parent
        I check the memory settings in ChatGPT every so often and it's never remembered anything for me. Maybe the questions I ask it are too impersonal?

        I check the memory settings in ChatGPT every so often and it's never remembered anything for me. Maybe the questions I ask it are too impersonal?

        1. [2]
          ackables
          Link Parent
          I’m on an edu license, so the memory may be better than free or personal licenses.

          I’m on an edu license, so the memory may be better than free or personal licenses.

          1 vote
          1. skybrian
            Link Parent
            Apparently I had no idea how memories work in ChatGPT. I tried these prompt instead: What information do you have about my chats over the past year? Did you need to ask for this log, or was it...

            Apparently I had no idea how memories work in ChatGPT. I tried these prompt instead:

            • What information do you have about my chats over the past year?

            • Did you need to ask for this log, or was it already provided?

            • What "user knowledge memories" do you have?

            The "saved memories" in settings is still blank and apparently doesn't work at all. So I guess the ChatGPT UI is not to be trusted and you should just ask questions instead? That's a new one.

  2. [4]
    skybrian
    Link
    I imagine that by "ad" they mean marketing, and that includes the cost of providing ChatGPT to free users. I don't know how to confirm that, though. I wonder what the non-cash operating losses are?

    OpenAI also spent US$2 billion on sales and ad

    I imagine that by "ad" they mean marketing, and that includes the cost of providing ChatGPT to free users. I don't know how to confirm that, though.

    Operating losses reached US$7.8 billion, and the company said it burned US$2.5 billion in cash.

    I wonder what the non-cash operating losses are?

    10 votes
    1. Lyrl
      Link Parent
      Maybe long payment terms to vendors? I know my employer has some materials they have negotiated 180 day payment terms with. Small startups generally don't have leverage to negotiate those kind of...

      Maybe long payment terms to vendors? I know my employer has some materials they have negotiated 180 day payment terms with. Small startups generally don't have leverage to negotiate those kind of terms, but these startups are not small.

      7 votes
    2. [2]
      stu2b50
      Link Parent
      I’d imagine mostly stock based compensation.

      I’d imagine mostly stock based compensation.

      1 vote
      1. skybrian
        Link Parent
        I wouldn't normally think of stock options as an operating loss, particularly not for people working in R&D. But maybe it would if they work in operations?

        I wouldn't normally think of stock options as an operating loss, particularly not for people working in R&D. But maybe it would if they work in operations?

        4 votes
  3. [6]
    CrypticCuriosity629
    Link
    I actually recently got rid of my ChatGPT subscription in favor of Claude because of how absolutely abysmal the accuracy has been with ChatGPT recently. I legitimately can't get straightforward or...

    I actually recently got rid of my ChatGPT subscription in favor of Claude because of how absolutely abysmal the accuracy has been with ChatGPT recently. I legitimately can't get straightforward or correct answers out of it anymore.

    9 votes
    1. [4]
      creesch
      Link Parent
      gpt-5 to me feels like an entirely different new model (family of models actually as there are multiple under the hood). Rather than a continuation, it has gotten much more verbose talking around...

      gpt-5 to me feels like an entirely different new model (family of models actually as there are multiple under the hood). Rather than a continuation, it has gotten much more verbose talking around the subject and direct question rather than about it.

      In the context of this post I wonder if the goal with gpt-5 is cost reduction rather than overal improvement.

      5 votes
      1. ibuprofen
        Link Parent
        I strongly agree with that. You can see the same thing happening in Gemini. The 03-25 experimental model was vastly superior to everything they've come out with since. I think they released an...

        In the context of this post I wonder if the goal with gpt-5 is cost reduction rather than overal improvement.

        I strongly agree with that.

        You can see the same thing happening in Gemini. The 03-25 experimental model was vastly superior to everything they've come out with since. I think they released an extremely impressive experimental model, got the accolades, and have spent the last 6 months on efficiency.

        2 votes
      2. [2]
        skybrian
        Link Parent
        I chose the "robot" personality and have "be brief" in my custom instructions. I haven't done an A-B test, but the verbosity seems okay. Maybe try that?

        I chose the "robot" personality and have "be brief" in my custom instructions. I haven't done an A-B test, but the verbosity seems okay. Maybe try that?

        1 vote
        1. creesch
          Link Parent
          Sure, I can adjust the base prompt and chose personalities (which all do the same thing behind the scenes). But that doesn't really change the fact that it really does feel like a different model...

          Sure, I can adjust the base prompt and chose personalities (which all do the same thing behind the scenes). But that doesn't really change the fact that it really does feel like a different model family. There are also some clear qualitative issues that can't really be fixed by tweaking the base prompt.

          It's why I said it likes to talk around subjects, rather than respond to the actual question. The various 4 models and o3 did a much better job of responding to the prompt in a way that GPT-5 simply doesn't.

          Which is fine, I am not really tied to chatGPT. My comment is more of an observation.

          2 votes
    2. Prodiggles
      Link Parent
      It's interesting because depending on the context of the question it's make or break. In the realm of computing, it's actually really good at getting me answers based on what is and what could be....

      It's interesting because depending on the context of the question it's make or break. In the realm of computing, it's actually really good at getting me answers based on what is and what could be.

      I have a new project I'm doing in the realm of developing that I wasn't sure was a novel type of scripting or not. I spent some time on an early build and had it analyze and look into similar or overlapping features. Discovered after doing other research it actually was novel and just needs to bake in some expected features. Getting a pros/cons and quick compare/contrast to see what things I should prioritize for adoption.

      This is likely for the fact computing and programming that documentation is king, so it makes it a lot easier finding what is real (API documentation) vs some random news cycle event and its effects (hardly documented, lots of opinions, not yet researched or baked in yet)

      4 votes