9 ms·

While technical AI and LLMs are not something I’m well versed in. So as I sit on the sidelines and see the current proliferation of AI startups I’m starting to

While technical AI and LLMs are not something I’m well versed in. So as I sit on the sidelines and see the current proliferation of AI startups I’m starting to wonder where the moats are outside of access to raw computing power. Open AI seemed to have a massive lead in this space but that lead seems to be shrinking every day.

weberer 2y ago
Obtaining high quality training data is the biggest moat right now.
- segasaturn 2y ago
  Where are they going to get that data? Everything on the open web after 2023 is polluted with lowquality AI slop that poisons the data sets. My prediction: Aggressive dragnet surveillance of users. As in, Google recording your phone calls on Android, Windows sending screen recordings from Recall to OpenAI, Meta training off Whatsapp messages... It sounds dystopian, but the Line Must Go Up.
  - jazzyjackson 2y ago
    I'm really curious if Microsoft will ever give in to the urge to train on private business data - since transitioning office to o365, they hold the world's and even governments word documents and emails. I'm pretty sure they've promised never to touch it but they can certainly read it so... Information wants to be free.
    - jhickok 2y ago
      Microsoft "trains" on business data already, but typically for things like fine-tuning security automation and recognizing malicious signals. It sure would be a big step to reading chats and email and feeding them in to a model.
    - ENGNR 2y ago
      Slack tried it but the backlash got them, this time anyway.
      - bossyTeacher 2y ago
        Elaborate? never heard of this
        ENGNR 2y ago
        El Reg got the story: https://www.theregister.com/2024/05/20/slack_ts_and_cs_update/ https://www.theregister.com/2024/05/20/slack_ts_and_cs_updat... They blamed it on a 'misunderstanding' in the end
    - wkat4242 2y ago
      Tbh it would also make the model much more accurate for corporate uses. It's not a bad idea for that reason. But security and DLP teams will never accept it.
  - crazygringo 2y ago
    > Everything on the open web after 2023 is polluted with lowquality AI slop that poisons the data sets. Not even close to everything. E.g. training on the NY Times and Wikipedia has zero meaningful AI. Training on books from reputable publishers similarly has zero meaningful AI. Any LLM usage was to polish prose or assist with research or whatever, but shouldn't affect the factual quality in any significant way. The web hasn't been polluted with AI any more than e-mail has been polluted with spam. Which is to say it's there, but it's also entirely viable to separate. Nobody's worried that the group email chain with friends is being overrun with spam or with AI.
    - dgfitz 2y ago
      [flagged]
    - mplewis 2y ago
      There is no heuristic that reliably separates human-written text from GPT text. It is impossible to avoid model collapse in such an environment.
      - ValentinA23 2y ago
        Unless you're OpenAI and provided like 98% of ai generated text that was posted to the internet.
      - Der_Einzige 2y ago
        Anti-slop sampler and bigram/trigram stylometric analysis absolutely can give evidence that someone used a particular LLM with particular settings. But in general it’s hard to separate human from AI text.
      - crazygringo 2y ago
        You're ignoring the information about the source. It's not about a heuristic on text of unknown provenance -- it's about publishers that exert a certain level of editorial control and quality verification. Or social reputation mechanisms that achieve the same. That's what is preventing your "model collapse". Reputations of provenance. Not pure-text heuristics.
  - lfmunoz4 2y ago
    Would think most quality data is books and news articles and scientific journals. Not crap people are texting each other. These companies will never admit it but AI is built on the back of piracy archives, easiest way and cheapest way to getting massive amounts of quality data.
    - mcmcmc 2y ago
      That entirely depends on what quality you’re going for. If the goal is to simulate passably human conversation, texts and dms are probably more desirable.
    - mswen 2y ago
      A friend and I built a proof-of-concept of using a variation of Latent Semantic Analysis to automatically build up conceptual maps and loadings of individual words against the latent conceptual vectors back in 2000. In exploring what it would take to scale I concluded, like you, that we should use professionally written and edited content like books, news articles and scientific journals as the corpus against which to build up the core knowledge graph. Twenty-four years later I still regret not being able to raise money to enable us to keep working on that nascent startup. In most ways it was still too early. Google was still burning through VC money at that point and the midwestern investors we had some access to didn't get it. And, honestly they were probably correct. Compute power was still too expensive and quality data sources like published text were mostly locked up and generally not available to harvest.
- staticautomatic 2y ago
  I’m in this space and no it isn’t.
  - sumedh 2y ago
    What is the moat then?
    - staticautomatic 2y ago
      Idk but it’s not lack of training data.
      - sumedh 2y ago
        You work in this space and you dont know what the moat is?
        staticautomatic 2y ago
        The market is two-sided
        sangnoir 2y ago
        All of the people working in CS don't know if P=NP: working in a field doesn't mean you have all the answers.
        mplewis 2y ago
        Tells you a lot about the state of the industry.
        tim333 2y ago
        There doesn't seem to be much of a moat. OpenAI, Gemini, Meta, X.Ai and Anthropic all seem to be able to much the same stuff. o1 is novel at the moment but I bet it'll be copied soon.
      - fastball 2y ago
        If you don't know what it is, how do you know it's not training data?
wongarsu 2y ago
Data. You want huge amounts of high quality data with a diverse range of topics, writing styles and languages. Everyone seems to balance those requirements a bit differently, and different actors have access to different training data There is also some moat in the refinement process (rlhf, model "safety" etc)
InkCanon 2y ago
You hit the nail on the head. Companies are scrambling for an edge. Not a real edge, an edge to convince investors to keep giving them money. Perplexity is going all in on convincing VCs it can create a "data flywheel".
- disqard 2y ago
  Perhaps I've missed something, but where will the infinite amounts of training data come from, for future improvements? If these models will be trained on the outputs of themselves (and other models), then it's not so much a "flywheel", as it is a Perpetual Motion Machine.
  - Tier3r 2y ago
    Perplexity has a dubious idea based around harvesting user chats -> making service better -> getting more user prompts. I am quite unconvinced that user prompts and stored chats will materially improve an LLM that is trained on a trillion high quality tokens. The second idea being kicked around is synthetic data will create a new fountain of youth for data that will also fix its reasoning abilities.
  - LarsDu88 2y ago
    There's pretraining which is just raw text from the internet but there's also supervised preference data sourced from users. Right now the edge is in acquiring the latter which OpenAI has a slight lead in
runeblaze 2y ago
In addition to data, having the infra to scale up training robustly is very very non-trivial.
YetAnotherNick 2y ago
> Open AI seemed to have a massive lead in this space but that lead seems to be shrinking every day. The lead is as strong as ever. They are 34 ELO above anyone else in blind testing, and 73 ELO above in coding [1]. They also seem to have artificially constrain the lead as they already have stronger model like o1 which they haven't released. Consistent to the past, they seem to release just <50 ELO above anyone else, and upgrades the model in weeks when someone gets closer. [1]: https://lmarena.ai/ https://lmarena.ai/
- adventured 2y ago
  It's rather amusing that people have said this about OpenAI - that they essentially had no lead - for about two years non-stop. The moat as usual is extraordinary scale, resources, time. Nobody is putting $10 billion into the 7th OpenAI clone. Big tech isn't aggressively partnering with the 7th OpenAI clone. The door is already shut to that 7th OpenAI clone (they can never succeed or catch-up), there's just an enormous amount of naivety in tech circles about how things work in the real world: I can just spin up a ChatGPT competitor over the weekend on my 5090, therefore OpenAI have no barriers to entry, etc. HN used to endlessly talk about how Uber could be cloned in a weekend. It's just people talking about something they don't actually understand. They might understand writing code (or similar) and their bias extends from the premise that their thing is the hard part of the equation (writing the code, building an app, is very far from the hardest part of the equation for an Uber).
  - TeaBrain 2y ago
    No-one was saying this 2 or even 1.5 years ago.
- epolanski 2y ago
  Idc about lmarena benchmarks, I test different models everyday in Cursor, Sonnet is way better at coding web applications than ChatGPT4o
  - Zetaphor 2y ago
    Completely agree. It's well known that the LMSys arena benchmarks are heavily skewed with bias towards whatever is new and exciting. Meanwhile even OpenAI have acknowledged Sonnet as being a superior coding model. This is clearly evident to anyone who spends any amount of time working on non-trivial projects with both models.
Der_Einzige 2y ago
How can anyone say that the lead is shrinking when no one still has any good competitor to strawberry? Dspy has been out for how long and how many folks have shown better reasoning models than strawberry built with literally anything else? Oh yeah, zero.
- mplewis 2y ago
  Wow, this thing can reason now? How come it keeps getting my basic word problems wrong?
  - Der_Einzige 2y ago
    Tokenization
    - youoy 2y ago
      Is it really a tested conclusion? Or a plausible conclusion to try to hide the limitations of the model architecture? I'm asking because I know that with some prompts it gets the answer correct, and in those cases nothing in the tokenization has changed.
      - Der_Einzige 2y ago
        Yes, this is 100% tested and proven ad nasum within the field. I have some of my own papers on this, but you can look at literally any major AI conference and find dozens of papers analyzing yet more issues caused by byte pair tokenization. Honestly the folks who don’t want to admit that it’s tokenization are just extremely salty that AI is actually good right now. Your “AI couldn’t tell me how many Rs in strawberry” stuff is extreme cope for your job prospects evaporating from a system that can’t spell correctly.
        youoy 2y ago
        But does a different prompt get the answer correct? I find it surprising. Can you share a link? I'm not saying this out of saltiness, I would be very grateful. If you don't want to I will try the shitty Google search, no problem.
_giorgio_ 2y ago
Quite the opposite. The new O1 models by open ai are surprisingly good. So good that you can see intelligence at work through solutions when you go from: GPT4 O1 mini O1 preview Way better than the competition. And they have a mobile voice app with advanced voice which is phenomenal.
PeterStuer 2y ago
Those extremely detailed models they can build of you (digital twin) from all your conversations coupled with intelligent subtle, likable, relatable and very personal artificial salesmen in service of the highest bidder has enourmous financial potential. Even if the latter becomes commoditized (and we are far from that in practice), the former is a serious mote. Just like their is no secret to building a search engine or a social network platform (and that is not saying their are no technical challanges), operating it profitably requires massive aggregate user profile exploitation potential which requires huge upfront loss leaders.
bboygravity 2y ago
The lead has been taken over by Xai already as far as I know. They seem to have 100k H100's up and running. OpenAI is not really leading the LLM world anyway ever since Claude sonnet 3.5 came out.
- belter 2y ago
  > OpenAI is not really leading the LLM world anyway ever since Claude sonnet 3.5 came out. Something many times repeated. It just takes a few minutes with the different models to find out it's not true.
  - csomar 2y ago
    I'd say it's 50/50 at the moment. Half the times, sonnet gives a definitive and better answer.
  - daghamm 2y ago
    I usually use them side by side, and very often (not always) Claude is better.
  - mnky9800n 2y ago
    I think the main issue with these metrics, which you implicitly highlight, Is that they are not a one size fits all approach. In fact, they are often treated, at least casually, like they are some kind of model fit like an r squared value. Which is maybe a good description narrowly constrained to the task or set of tasks they are being evaluated on for the metric. But the complexity of the user experience combined with the poor sample rate that a person can individually experience leads to conclusions like these. And they are perfectly valid conclusions. If the model doesn’t work for you, why use it? But it also suggests that personal experience cannot be used to decide if the model performs in aggregate well or not. But this doesn’t matter to the individual user or problem space. Because they should of course use whatever works best for them.
browningstreet 2y ago
OpenAI just raised a huge pile of money. I’m sure the most was a consideration during the funding raise. I’m guessing the issues between openAI and Ms are beyond just the moat and competitiveness’s. Every OpenAI thread focuses on the moat, but surely that was baked into their business dealings of the last 60 days.
- appplication 2y ago
  > I’m sure the most was a consideration during the funding raise Ha. Really though, the entirety of venture capital could be summarized as “this probably won’t pay out but if it does it’s gonna be epic”. I wouldn’t read too much into America’s thirstiest capitalists spending their discretionary billions on the latest hype cycle.

weberer 2y ago

Obtaining high quality training data is the biggest moat right now.

segasaturn 2y ago
Where are they going to get that data? Everything on the open web after 2023 is polluted with lowquality AI slop that poisons the data sets. My prediction: Aggressive dragnet surveillance of users. As in, Google recording your phone calls on Android, Windows sending screen recordings from Recall to OpenAI, Meta training off Whatsapp messages... It sounds dystopian, but the Line Must Go Up.
- jazzyjackson 2y ago
  I'm really curious if Microsoft will ever give in to the urge to train on private business data - since transitioning office to o365, they hold the world's and even governments word documents and emails. I'm pretty sure they've promised never to touch it but they can certainly read it so... Information wants to be free.
  - jhickok 2y ago
    Microsoft "trains" on business data already, but typically for things like fine-tuning security automation and recognizing malicious signals. It sure would be a big step to reading chats and email and feeding them in to a model.
  - ENGNR 2y ago
    Slack tried it but the backlash got them, this time anyway.
    - bossyTeacher 2y ago
      Elaborate? never heard of this
      - ENGNR 2y ago
        El Reg got the story: https://www.theregister.com/2024/05/20/slack_ts_and_cs_update/ https://www.theregister.com/2024/05/20/slack_ts_and_cs_updat... They blamed it on a 'misunderstanding' in the end
  - wkat4242 2y ago
    Tbh it would also make the model much more accurate for corporate uses. It's not a bad idea for that reason. But security and DLP teams will never accept it.
- crazygringo 2y ago
  > Everything on the open web after 2023 is polluted with lowquality AI slop that poisons the data sets. Not even close to everything. E.g. training on the NY Times and Wikipedia has zero meaningful AI. Training on books from reputable publishers similarly has zero meaningful AI. Any LLM usage was to polish prose or assist with research or whatever, but shouldn't affect the factual quality in any significant way. The web hasn't been polluted with AI any more than e-mail has been polluted with spam. Which is to say it's there, but it's also entirely viable to separate. Nobody's worried that the group email chain with friends is being overrun with spam or with AI.
  - dgfitz 2y ago
    [flagged]
  - mplewis 2y ago
    There is no heuristic that reliably separates human-written text from GPT text. It is impossible to avoid model collapse in such an environment.
    - ValentinA23 2y ago
      Unless you're OpenAI and provided like 98% of ai generated text that was posted to the internet.
    - Der_Einzige 2y ago
      Anti-slop sampler and bigram/trigram stylometric analysis absolutely can give evidence that someone used a particular LLM with particular settings. But in general it’s hard to separate human from AI text.
    - crazygringo 2y ago
      You're ignoring the information about the source. It's not about a heuristic on text of unknown provenance -- it's about publishers that exert a certain level of editorial control and quality verification. Or social reputation mechanisms that achieve the same. That's what is preventing your "model collapse". Reputations of provenance. Not pure-text heuristics.
- lfmunoz4 2y ago
  Would think most quality data is books and news articles and scientific journals. Not crap people are texting each other. These companies will never admit it but AI is built on the back of piracy archives, easiest way and cheapest way to getting massive amounts of quality data.
  - mcmcmc 2y ago
    That entirely depends on what quality you’re going for. If the goal is to simulate passably human conversation, texts and dms are probably more desirable.
  - mswen 2y ago
    A friend and I built a proof-of-concept of using a variation of Latent Semantic Analysis to automatically build up conceptual maps and loadings of individual words against the latent conceptual vectors back in 2000. In exploring what it would take to scale I concluded, like you, that we should use professionally written and edited content like books, news articles and scientific journals as the corpus against which to build up the core knowledge graph. Twenty-four years later I still regret not being able to raise money to enable us to keep working on that nascent startup. In most ways it was still too early. Google was still burning through VC money at that point and the midwestern investors we had some access to didn't get it. And, honestly they were probably correct. Compute power was still too expensive and quality data sources like published text were mostly locked up and generally not available to harvest.
staticautomatic 2y ago
I’m in this space and no it isn’t.
- sumedh 2y ago
  What is the moat then?
  - staticautomatic 2y ago
    Idk but it’s not lack of training data.
    - sumedh 2y ago
      You work in this space and you dont know what the moat is?
      - staticautomatic 2y ago
        The market is two-sided
      - sangnoir 2y ago
        All of the people working in CS don't know if P=NP: working in a field doesn't mean you have all the answers.
      - mplewis 2y ago
        Tells you a lot about the state of the industry.
      - tim333 2y ago
        There doesn't seem to be much of a moat. OpenAI, Gemini, Meta, X.Ai and Anthropic all seem to be able to much the same stuff. o1 is novel at the moment but I bet it'll be copied soon.
    - fastball 2y ago
      If you don't know what it is, how do you know it's not training data?

wongarsu 2y ago

Data. You want huge amounts of high quality data with a diverse range of topics, writing styles and languages. Everyone seems to balance those requirements a bit differently, and different actors have access to different training data There is also some moat in the refinement process (rlhf, model "safety" etc)

InkCanon 2y ago

You hit the nail on the head. Companies are scrambling for an edge. Not a real edge, an edge to convince investors to keep giving them money. Perplexity is going all in on convincing VCs it can create a "data flywheel".

disqard 2y ago
Perhaps I've missed something, but where will the infinite amounts of training data come from, for future improvements? If these models will be trained on the outputs of themselves (and other models), then it's not so much a "flywheel", as it is a Perpetual Motion Machine.
- Tier3r 2y ago
  Perplexity has a dubious idea based around harvesting user chats -> making service better -> getting more user prompts. I am quite unconvinced that user prompts and stored chats will materially improve an LLM that is trained on a trillion high quality tokens. The second idea being kicked around is synthetic data will create a new fountain of youth for data that will also fix its reasoning abilities.
- LarsDu88 2y ago
  There's pretraining which is just raw text from the internet but there's also supervised preference data sourced from users. Right now the edge is in acquiring the latter which OpenAI has a slight lead in

runeblaze 2y ago

In addition to data, having the infra to scale up training robustly is very very non-trivial.

YetAnotherNick 2y ago

> Open AI seemed to have a massive lead in this space but that lead seems to be shrinking every day. The lead is as strong as ever. They are 34 ELO above anyone else in blind testing, and 73 ELO above in coding [1]. They also seem to have artificially constrain the lead as they already have stronger model like o1 which they haven't released. Consistent to the past, they seem to release just <50 ELO above anyone else, and upgrades the model in weeks when someone gets closer. [1]: https://lmarena.ai/ https://lmarena.ai/

adventured 2y ago
It's rather amusing that people have said this about OpenAI - that they essentially had no lead - for about two years non-stop. The moat as usual is extraordinary scale, resources, time. Nobody is putting $10 billion into the 7th OpenAI clone. Big tech isn't aggressively partnering with the 7th OpenAI clone. The door is already shut to that 7th OpenAI clone (they can never succeed or catch-up), there's just an enormous amount of naivety in tech circles about how things work in the real world: I can just spin up a ChatGPT competitor over the weekend on my 5090, therefore OpenAI have no barriers to entry, etc. HN used to endlessly talk about how Uber could be cloned in a weekend. It's just people talking about something they don't actually understand. They might understand writing code (or similar) and their bias extends from the premise that their thing is the hard part of the equation (writing the code, building an app, is very far from the hardest part of the equation for an Uber).
- TeaBrain 2y ago
  No-one was saying this 2 or even 1.5 years ago.
epolanski 2y ago
Idc about lmarena benchmarks, I test different models everyday in Cursor, Sonnet is way better at coding web applications than ChatGPT4o
- Zetaphor 2y ago
  Completely agree. It's well known that the LMSys arena benchmarks are heavily skewed with bias towards whatever is new and exciting. Meanwhile even OpenAI have acknowledged Sonnet as being a superior coding model. This is clearly evident to anyone who spends any amount of time working on non-trivial projects with both models.

Der_Einzige 2y ago

How can anyone say that the lead is shrinking when no one still has any good competitor to strawberry? Dspy has been out for how long and how many folks have shown better reasoning models than strawberry built with literally anything else? Oh yeah, zero.

mplewis 2y ago
Wow, this thing can reason now? How come it keeps getting my basic word problems wrong?
- Der_Einzige 2y ago
  Tokenization
  - youoy 2y ago
    Is it really a tested conclusion? Or a plausible conclusion to try to hide the limitations of the model architecture? I'm asking because I know that with some prompts it gets the answer correct, and in those cases nothing in the tokenization has changed.
    - Der_Einzige 2y ago
      Yes, this is 100% tested and proven ad nasum within the field. I have some of my own papers on this, but you can look at literally any major AI conference and find dozens of papers analyzing yet more issues caused by byte pair tokenization. Honestly the folks who don’t want to admit that it’s tokenization are just extremely salty that AI is actually good right now. Your “AI couldn’t tell me how many Rs in strawberry” stuff is extreme cope for your job prospects evaporating from a system that can’t spell correctly.
      - youoy 2y ago
        But does a different prompt get the answer correct? I find it surprising. Can you share a link? I'm not saying this out of saltiness, I would be very grateful. If you don't want to I will try the shitty Google search, no problem.

_giorgio_ 2y ago

Quite the opposite. The new O1 models by open ai are surprisingly good. So good that you can see intelligence at work through solutions when you go from: GPT4 O1 mini O1 preview Way better than the competition. And they have a mobile voice app with advanced voice which is phenomenal.

PeterStuer 2y ago

Those extremely detailed models they can build of you (digital twin) from all your conversations coupled with intelligent subtle, likable, relatable and very personal artificial salesmen in service of the highest bidder has enourmous financial potential. Even if the latter becomes commoditized (and we are far from that in practice), the former is a serious mote. Just like their is no secret to building a search engine or a social network platform (and that is not saying their are no technical challanges), operating it profitably requires massive aggregate user profile exploitation potential which requires huge upfront loss leaders.

bboygravity 2y ago

The lead has been taken over by Xai already as far as I know. They seem to have 100k H100's up and running. OpenAI is not really leading the LLM world anyway ever since Claude sonnet 3.5 came out.

belter 2y ago
> OpenAI is not really leading the LLM world anyway ever since Claude sonnet 3.5 came out. Something many times repeated. It just takes a few minutes with the different models to find out it's not true.
- csomar 2y ago
  I'd say it's 50/50 at the moment. Half the times, sonnet gives a definitive and better answer.
- daghamm 2y ago
  I usually use them side by side, and very often (not always) Claude is better.
- mnky9800n 2y ago
  I think the main issue with these metrics, which you implicitly highlight, Is that they are not a one size fits all approach. In fact, they are often treated, at least casually, like they are some kind of model fit like an r squared value. Which is maybe a good description narrowly constrained to the task or set of tasks they are being evaluated on for the metric. But the complexity of the user experience combined with the poor sample rate that a person can individually experience leads to conclusions like these. And they are perfectly valid conclusions. If the model doesn’t work for you, why use it? But it also suggests that personal experience cannot be used to decide if the model performs in aggregate well or not. But this doesn’t matter to the individual user or problem space. Because they should of course use whatever works best for them.

browningstreet 2y ago

OpenAI just raised a huge pile of money. I’m sure the most was a consideration during the funding raise. I’m guessing the issues between openAI and Ms are beyond just the moat and competitiveness’s. Every OpenAI thread focuses on the moat, but surely that was baked into their business dealings of the last 60 days.

appplication 2y ago
> I’m sure the most was a consideration during the funding raise Ha. Really though, the entirety of venture capital could be summarized as “this probably won’t pay out but if it does it’s gonna be epic”. I wouldn’t read too much into America’s thirstiest capitalists spending their discretionary billions on the latest hype cycle.