[HN Gopher] Things I don't know about AI
       ___________________________________________________________________
        
       Things I don't know about AI
        
       Author : todsacerdoti
       Score  : 155 points
       Date   : 2024-02-21 13:40 UTC (9 hours ago)
        
 (HTM) web link (blog.eladgil.com)
 (TXT) w3m dump (blog.eladgil.com)
        
       | eladgil wrote:
       | Would love to get any views / counterpoints on these topics.
       | 
       | Eg which AI infra approach thrives relative to the big clouds?
       | 
       | What new architectures are needed for agents vs transformers?
       | 
       | Etc...
       | 
       | Lots of stuff to figure out....
        
         | texas2toss wrote:
         | Do you think there is the possibility of consumer or end-user
         | apps collecting enough specialized data to move downwards on
         | your graph to infra and foundational models?
        
           | eladgil wrote:
           | I think this is already true for things based on diffusion
           | models - e.g. Midjourney, Pika etc
           | 
           | For LLMs, Character and ChatGPT are arguably two vertically
           | integrated consumer apps (with some B2B applications for
           | ChatGPT as well)
        
         | danielmarkbruce wrote:
         | There is a decent (<50%, >20%) chance that frontier foundation
         | models are less oligopoly like than it seems. The reason is
         | that there are so many levers to pull, so much low hanging
         | fruit.
         | 
         | For example: * Read the Bloomberg GPT paper - they create their
         | own tokenizer. For specialized domains (finance, law, medicine,
         | etc) the vocabulary is very different and there is likely a lot
         | to do here, where individual tokens really need to map to
         | specific concepts and having a concept capture in several
         | tokens makes it too hard to learn on limited domain data. *
         | Data - so many ways to do different data - more/less, cleaner,
         | "better" on some dimension. * Read the recent papers on
         | different decoding strategies - there seems to be a lot to do
         | here. * Model architecture (SSM etc). If you speak to people
         | who aren't even researchers, they have 10 ideas around
         | architecture and some of them are decent sounding ideas - lots
         | of low hanging fruit. * System architecture - ie likely to see
         | more and more "models" served via API which are actually
         | systems of several model calls, and there is a lot to do here.
         | * Hardware, lower precision etc likely to make training much
         | cheaper
         | 
         | It's reasonably likely (again, guessing < 50% > 20%) that this
         | large set of levers to pull become ways to see constant leap-
         | frogging for years and years. Or, at least they become
         | choices/trade-offs rather than strictly "better".
        
           | eladgil wrote:
           | I agree this is a potential outcome. One big question is
           | generalizability versus niche models. For example, is the
           | best legal model a frontier model + a giant context window +
           | RAG? Or is it a niche model trained or fine tuned for law?
           | 
           | Right now at least people seem to decouple some measures of
           | how smart the model is from knowledge base, and at least for
           | now the really big models seem smartest. So part of the
           | question is well is how insightful / synthesis centric the
           | model needs to be versus effectively doing regressions....
        
             | danielmarkbruce wrote:
             | Yup, it's unclear. The current ~consensus is "general
             | purpose frontier model + very sophisticated RAG/system
             | architecture" for legal as an example. I'm building
             | something here using this idea and think its 50/50 (at
             | best) I'm on the right path. It's quite easy to build very
             | clever sounding but often wrong insights into various legal
             | agreements (m&a docs for example). When looking at the
             | tokenization, the training data, decode, architecture (lots
             | of guesses) of the big models, there are a lot of things
             | where the knobs seem turned slightly incorrectly for the
             | domain.
             | 
             | Some of the domains are so large that a specialized model
             | might seem niche but the value prop is potentially
             | astronomical.
        
             | CuriouslyC wrote:
             | Frontier model + rag is good when you need cross-discipline
             | abilities and general knowledge, niche models are best when
             | the domain is somewhat self contained (for instance, if you
             | wanted a model that is amazing at role playing certain
             | types of characters).
             | 
             | The future is model graphs with networked mixtures of
             | experts, where models know about other models and can call
             | them as part of recursive prompts, with some sort of online
             | training to tune the weights of the model graph.
        
               | sfink wrote:
               | > The future is model graphs with networked mixtures of
               | experts, where models know about other models and can
               | call them as part of recursive prompts, with some sort of
               | online training to tune the weights of the model graph.
               | 
               | What's the difference between that and combining all of
               | the models into a single model? Aren't you just
               | introducing limitations in communication and training
               | between different parts of that uber-model, limitations
               | that may as well be encoded into the single model if
               | they're useful? Are you just partitioning for training
               | performance? Which is a big deal, of course, but it just
               | seems like guessing the right partitioning and
               | communication limitations is not going to be
               | straightforward compared to the usual stupid "throw it
               | all in one big pile and let it work itself out" approach.
        
               | CuriouslyC wrote:
               | The limitation is the amount of model you can fit on your
               | hardware, and also sometimes information about one domain
               | can incorrectly introduce biases in another which are
               | very hard to fix, so training on one domain only will
               | produce much better results.
        
         | jchonphoenix wrote:
         | Arguably one of the earliest consumer use cases that found
         | footing was AI girlfriend/boyfriend. Large amounts of revenue
         | spread across many small players are generated here but it's
         | glossed over due to the category.
        
           | notpachet wrote:
           | I think one related area we'll start seeing more of in the
           | future is "resurrected" companions. You have a terminally ill
           | family member, so you train a model on a bunch of video
           | recordings of them, then you can talk to "them" after they've
           | shuffled off this mortal coil.
        
             | swozey wrote:
             | Licensing my soul as closed source now ..
        
               | datadrivenangel wrote:
               | DNR: Do Not Retrain
        
             | babyshake wrote:
             | Be right back.
        
           | rockostrich wrote:
           | Weird how close your user name is to Joaquin Phoenix, star of
           | the film "Her" centered around an AI girlfriend.
        
         | doctorpangloss wrote:
         | I haven't listen to your great podcasts so hard to say what is
         | not covered.
         | 
         | Architectures matter a lot less than data. "Knowledge" and
         | "reasoning" in LLMs is a manifestation of instruct-style data.
         | It won't matter how much cheaper training gets if there is
         | limited instruct data for use cases.
         | 
         | How do you make 100k context window data for example? Still
         | need thousands of people. Same with so-called niches.
         | 
         | Maybe it turns out to be a complex coordination problem to
         | share the data. That's bad for equity investors and giant
         | companies. Anyway, all of this would cost less than the moon
         | landing so it's practicable, you don't need cheaper, you just
         | need risk-taking.
         | 
         | The obviousness of the path from here to there means it's not
         | about innovation. It's all about strategy.
         | 
         | If Google could marshal $10b for Stadia it could spend $10b on
         | generating 100k context window instruct style data and have the
         | best model. It could also synthesize videos from Unity/Unreal
         | for Sora-style generation. It would just be very hard in an org
         | with 100,000+ people to spend $10b on 10 developers and 10,000
         | writers compared to 400 developers and 3,600 product managers
         | and other egos. At the end of the day you are revisiting the
         | weaknesses that brought Google and other big companies to this
         | mess in the first place.
         | 
         | Anyway I personally think the biggest weakness with ChatGPT and
         | the chat-style UX is that it feels like work. Netflix, TikTok,
         | etc. don't feel like work. Nobody at Google (or OpenAI for that
         | matter) knows how to make stuff that doesn't feel like work.
         | And you can't measure "fun." So the biggest thing to figure out
         | is how much technical stuff matters in a world where people can
         | be poached here and there and walk out with the whole
         | architecture in their heads, versus the non-technical stuff
         | that takes decades of hard-worn personal experiences and
         | strongly held opinions like answers to questions "How do you
         | make AI fun?"
        
           | mywittyname wrote:
           | > "How do you make AI fun?"
           | 
           | Bring back text-based adventure games.
        
             | doctorpangloss wrote:
             | People go for this level of obviousness and it doesn't
             | work. I have no doubt that a meme text game will find some
             | level of literal objective success. But it will still suck.
             | Meme games are a terrible business, both in terms of
             | profits and equity.
             | 
             | This also speaks to why OpenAI, Google and the other
             | developers will struggle to create anything that feels like
             | fun: they will chase obvious stuff like this, they will
             | think its similar to all problems. And in reality, you
             | don't need any testing or data or whatever to know that
             | people hate reading in video games, the best video game
             | writing is worse than the average movie's screenplay, that
             | most dialogue is extremely tedious, so why are you going to
             | try to make it even worse by making it generated by an AI?
        
       | xianshou wrote:
       | From the article:
       | 
       | "It is important to note that the scale of investments being made
       | by these cloud providers is dwarfed by actual cloud revenue. For
       | example, Azure from Microsoft generates $25B in revenue a
       | quarter. The ~$10B OpenAI investment by Microsoft is roughly 6
       | weeks of Azure revenue. This suggests the cloud business (at
       | least for now) is more important than any one model set for
       | Azure."
       | 
       | The most interesting implication in the short term is that the
       | impact of model availability on cloud provider choice is more
       | important than the models themselves. Will organizations choose
       | Azure because it offers GPT-4, or AWS because it offers Claude?
       | 
       | This also explains reciprocal VC structures whereby e.g. Amazon
       | invests billions in Anthropic while Anthropic in turn promises to
       | spend billions on AWS cloud resources. Anthropic gets a higher
       | valuation and broader distribution, while AWS gets more
       | customers. It would be a fascinating outcome if the predicted
       | oligopoly in fact ends up as several exclusive partnerships.
        
         | infecto wrote:
         | Doubtful, at the end of the day the models are just like any
         | other API. No reason not to signup for all the big players
         | (GCP/Azure/AWS) and have access to the various models. Should
         | have no weight as to where you should build out your infra.
        
         | simpletone wrote:
         | > The ~$10B OpenAI investment by Microsoft is roughly 6 weeks
         | of Azure revenue.
         | 
         | Revenue isn't what is important. Profit is since profit is the
         | money you have after costs to invest. But since cloud is highly
         | profitable as it is, the point still stands I suppose.
         | 
         | > It would be a fascinating outcome if the predicted oligopoly
         | in fact ends up as several exclusive partnerships.
         | 
         | I thought people were predicting independent silos? Where
         | highly profitable and cash rich companies like apple,
         | microsoft, google, facebook, etc own their entire stack or
         | exist as their own ecosystem. Where the likes of apple does
         | everything in-house - AI, cloud, OS, chips, etc. The whole she-
         | bang.
        
           | bbor wrote:
           | Opened the comments for exactly this. This person seems very
           | nice but also convinced that they're a business maven, which
           | I think leads them to making "clear" observations like that.
           | Microsoft hasn't invested a full _year_ of _revenue_ into
           | OpenAI, so it's not that invested? Meh
        
       | bane wrote:
       | The confusion that everybody feels right now about "what do I do
       | in this rapidly evolving space" is likely more an artifact of one
       | of the most vertical lead-edges to a hype curve in a generation.
       | There is so much piling on, at absolutely tremendous investment
       | amounts, that it's not really even possible to pick a launch
       | point to a business plan that would survive even months at this
       | point.
       | 
       | The second problem is more obvious, but not talked about in
       | concrete terms yet. As these technologies mature, the legal,
       | policy, and social constructs around them will also mature, and
       | will shape the future of what is coming as much as any technical
       | or business decision. Some examples:
       | 
       | - Imagine if courts _really_ find that these models are in gross
       | copyright violation and that model builders will need to work
       | through a scheme similar to sampling in music in order to do
       | their work into the future. Not only will it dramatically
       | increase the cost of model development, but will drive some of
       | the work to alternative sources of data if it 's cheaper. What
       | effect will this have on these models?
       | 
       | - Suppose these models are found to be reliable enough for
       | certain mission critical work (healthcare, defense, etc.) and
       | they lead a user down a path that results in harm or death? Who's
       | at fault? Suppose it becomes a liability of the company offering
       | the model? How does the insurance market respond?
       | 
       | - What if these models become good enough, and the automation
       | around them as well, that we start turning them into fully
       | automated agents that make up a significant part of our economy?
       | Do people still have jobs? Or will these things just become the
       | equivalent of hiring more people into an economy and we work in a
       | blended AI/Human Intelligence world? What happens when these
       | things decide to tank capital markets or choke supply chains?
       | 
       | - What about the long-term use of these models by individuals
       | with severe mental impairments or psychological issues? What
       | happens when those models go down, or produce a response that's
       | not liked, or there's a loss of historical conversational data?
       | What about otherwise normal people who end up down a path of
       | social isolation?
       | 
       | - There's an entire legal market about to be created for remote-
       | work fraud where somebody just replaces themself with one of
       | these models for 90% of their work. It's sci-fi utopia, but
       | fraudulent if you misrepresent who's doing the work in many
       | cases.
       | 
       | - Some company, somewhere, will crack the nut on figuring out how
       | to improve the speed and performance of these models, and build
       | custom hardware that offloads the entire effort onto rapidly
       | commoditizing systems. When multi-modal LLMs and diffusion models
       | come preloaded on disposable $.10 SOCs that are put on cereal
       | boxes and in Birthday cards, we'll be well beyond every one of
       | these questions. But we'll be firmly in a different future.
       | 
       | We have yet to encounter most of these things to the point that
       | as a society we have to build responses, and that itself points
       | to the immaturity of both these technologies and the use cases
       | for which they will be adopted.
       | 
       | Which cloud to pick and why aren't consumers spending thousands
       | of dollars getting the complex systems in place to use this stuff
       | aren't even the problems that are going to become interesting in
       | the next 10 years.
        
         | xamuel wrote:
         | >Suppose these models are found to be reliable enough for
         | certain mission critical work (healthcare, defense, etc.) and
         | they lead a user down a path that results in harm or death?
         | Who's at fault? Suppose it becomes a liability of the company
         | offering the model? How does the insurance market respond?
         | 
         | So much this. I think a lot of people overlook this aspect so
         | much. I wonder if we'll see a corporate push to make people
         | think of individual instantiations of LLMs as being their own
         | entities, for scapegoating purposes. It's completely ludicrous,
         | but you know they'd just love to point a finger at Machine
         | #73583 and make that take all the blame.
        
           | chasd00 wrote:
           | > corporate push to make people think of individual
           | instantiations of LLMs as being their own entities, for
           | scapegoating purposes
           | 
           | an airline already tried this.
           | 
           | "Air Canada argues it cannot be held liable for information
           | provided by one of its agents, servants, or representatives -
           | including a chatbot. It does not explain why it believes that
           | is the case. In effect, Air Canada suggests the chatbot is a
           | separate legal entity that is responsible for its own
           | actions."
           | 
           | https://www.forbes.com/sites/marisagarcia/2024/02/19/what-
           | ai...
        
         | shermantanktop wrote:
         | What happens when AI-generated content organically becomes a
         | large percentage of the user-generated corpus that is used for
         | training, with no way to differentiate it from human-generated
         | content?
         | 
         | We may look back on 2022/3 as the last time we had a training
         | set that was clean. The fact that it was already polluted with
         | SEO garbage will be a quaint problem.
        
           | bane wrote:
           | The current thinking is that it becomes a negative feedback
           | loop, a kind of lossy "compression" of our cultural output.
           | Combined with normal bitrot we'll eventually start losing
           | components of our shared culture until civilization ends.
        
         | somewhereoutth wrote:
         | - What happens if everyone realises that the areas where LLMs
         | can be usefully applied are few and not particularly lucrative?
        
           | bane wrote:
           | That's the far right of the hype curve.
        
       | cs702 wrote:
       | I would add to the list a few questions about the evolution of
       | cost dynamics going forward, given the advent of new sequence
       | modeling architectures.
       | 
       | As we all know, Transformers are _very_ expensive to train and
       | run because their compute cost is _quadratic_ in context length:
       | O(N2).
       | 
       | If newer model architectures like RWKV, Mamba, and various
       | others, which incur cost that is _linear_ in context length,
       | O(N), prove as successful as Transformers, the demand for compute
       | for a query of N tokens would decline from O(N2) to O(N).
       | 
       | For a sequence with N = 100,000 tokens, it would mean _cost
       | dropping by a factor of 100,000x_. That 's not peanuts!
       | 
       | The implications for all market participants would be
       | significant.
        
         | nyrikki wrote:
         | In the general case that would be the same as hierarchy
         | collapse of PH, which would prove P=NP
        
           | cs702 wrote:
           | Well, it would depend on how many tokens from the vocabulary
           | are required, in the worst case, to solve NP problems.
           | 
           | Think of the tokens as symbols, and of the vocabulary as a
           | finite alphabet of symbols, in a formal system.
           | 
           | In other words, no, it would NOT necessarily prove P = NP.
        
             | nyrikki wrote:
             | The Singleton case is NP-complete
             | 
             | If you consider NP second-order existential logic, meaning
             | that given x, you will get a y, While there may be many
             | true values that aren't y, it will help
             | 
             | Attention, which requires feed forward networks, which
             | approximate a DAG, you can think of attention as run time
             | reweighing to change what y value is returned.
             | 
             | PH is a query expressible in second-order logic, which
             | would allow for 'for any' at the second level.
             | 
             | The reduction in the general case to just logarithmic time:
             | O(n log n), would cause issue, there will not be any
             | movement from quadratic time to linear time without massive
             | new discoveries.
             | 
             | It is inherent from the restrictions of ANNs having binary
             | output.
             | 
             | Considering FP and FNP vs P and NP may help as they make
             | that property clearer than the concept of second-order
             | existential logic does in some contexts.
        
               | cs702 wrote:
               | Thank you. Right now, my puny little brain is searching
               | through what I would describe as "those distant and
               | slightly vague memories" that make up my poor man's
               | knowledge of computational complexity theory.
               | 
               |  _> Attention, which requires feed forward networks,
               | which approximate a DAG, you can think of attention as
               | run time reweighing to change what y value is returned._
               | 
               | Isn't that's true for linear RNN mechanisms too? Note
               | that linear RNNs like RWKV and Mamba are linear in the
               | number of tokens in the sequence, but not in the number
               | of features per token.
        
               | nyrikki wrote:
               | You don't typically get to the Sigma_2^P level unless are
               | a sadist or forced.
               | 
               | Attention works differently in RNNs, and it may be able
               | to find different tractable forms, and may solve
               | different problems, but increasing generalization is a
               | reduction of compression in the general case.
               | 
               | A many to one reduction to a recursively enumerable set
               | gets you to finite time, a reduction to NP gets you to
               | quadratic time.
               | 
               | Obviously those are upper limits and specific instances
               | or classes of instances may do better. P being inside NP
               | is a good example.
               | 
               | All ANNs are binary output, which helps with that. SNN or
               | spikey neutral networks, which better model cortical
               | neurons and have continuous output have problems with
               | being un-computable without much complexity as an
               | example.
               | 
               | That can be viewed as the problems with computable
               | numbers in place of the reals. But there are lots of
               | interesting and practical things to solve. It is just
               | those solutions will be more domain specific IMHO.
        
           | dadadad100 wrote:
           | As in the case of constraint solvers there are tons of
           | heuristics that have been applied over the years to make
           | performance look like it scales somewhat linearly in many or
           | most problems. This doesn't mean p==np
        
             | nyrikki wrote:
             | Most of those have had practical deciders functions.
             | 
             | If you consider NP-complete as the intersection of NP,
             | which is a decision problem, and NP-hard which a decider
             | function isn't possible; look at the success of heuristics
             | between those two sets.
             | 
             | Existential quantifiers on the second order terms in NP is
             | the same thing.
             | 
             | Over Parameterization can be thought as less lossy
             | compression and attention can be thought of as improved
             | retrieval or as additional expressiveness by not missing
             | out on as many outputs using binary output.
             | 
             | You can consider how LLMs tend to append, resulting in
             | telescope reduction as one of those approximation
             | reductions.
             | 
             | But this is due to the combinatorial optimization, which is
             | far more difficult.
             | 
             | Another way to think about it is that PH is equal to the
             | set of boolean queries on a concurent random acess machine
             | using exponentially many processors and constant time.
             | 
             | If your datasets have the Markovian property and are close
             | to ergotic there are options, but you probably wouldn't
             | need to resort to attention and over parameterization in
             | that case.
             | 
             | The word sense disambiguation problem is probably another
             | lens. That is harder yet but may be a way to think about
             | what I am trying to explain.
        
               | dadadad100 wrote:
               | Thanks for this reply. HN paid for itself again today
        
         | doctorpangloss wrote:
         | > The implications [of a bunch of different architectures] for
         | all market participants would be significant.
         | 
         | I don't know, who cares. How much good training data is there
         | that correctly exercises 100k context windows?
         | 
         | All the costs and complexity are tied up in authoring good
         | training data, not compute. 1 person can invent an alternative
         | easier to train architecture. 1 person cannot author 100k
         | context window instruct examples with the same quality as the
         | 4k ones. It will still take thousands of people for 100k just
         | like it took 1000s of people for 4k windows, image labeling,
         | etc.
        
           | og_kalu wrote:
           | >All the costs and complexity are tied up in authoring good
           | training data, not compute.
           | 
           | No it's compute. Post-Training by Human reinforcement is not
           | necessary. Anthropic employs RLAIF and it works just fine.
           | Some don't bother with reinforcement learning at all and just
           | leave it at fine-tuning on Instruction-response pairs.
           | 
           | The work being done to pre or post-training data is
           | insignificant in comparison.
           | 
           | You don't need 100k instruct-tuning examples. The vast
           | majority of instruct-tuning data is nowhere maxing even a 4k
           | context.
           | 
           | 100k _pre-training_ runs would probably be very helpful but
           | the thing stopping that from happening even in domains that
           | regularly match or exceed that context (fiction, law, code,
           | etc) is the ridiculous compute it would require to train with
           | that much context.
        
             | doctorpangloss wrote:
             | > The vast majority of instruct-tuning data is nowhere
             | maxing even a 4k context
             | 
             | This should indicate for you how much the problem really is
             | the training data.
        
               | og_kalu wrote:
               | It shows you don't need a instruction-tuning dataset that
               | wide to create a model that follows instructions for
               | large contexts.
               | 
               | I don't know how else to tell you that you're completely
               | wrong on what costs construe the majority of LLM
               | development/training.
        
         | araes wrote:
         | With the Big O notation, are there also possibilities for
         | either the whole process to go to O(nlogn) or possibly
         | subportions to the O(logn) range? There's also a lot of math
         | operations over the years that have gone to O(Na) where 1 < a <
         | 2.
        
         | jxy wrote:
         | > architectures like RWKV, Mamba, and various others, which
         | incur cost that is linear in context length, O(N)
         | 
         | These are actually O(N*M) where M is independent of inference
         | time context length, but it controls the actual amount of
         | information that gets retained during inference. Ignoring that
         | factor of M is misleading.
        
           | hansvm wrote:
           | All models are wrong. Some are useful.
           | 
           | The "linear" models I'm aware of have a cost of O(N * M),
           | where M is the amount of information retaind _per token_
           | (with some exceptions that store less state but still have
           | intermediate computations for some M about the same size as
           | the others, and with less state you necessarily have
           | forgetting issues and can't scale to unbounded sequences
           | anyway).
           | 
           | Contrast that with a transformer, which has a cost of O(N^2 *
           | M), where M is still the amount of information retained _per
           | token_. Dividing the two you still find a multiplicative
           | speedup of O(N).
           | 
           | For accurate cost estimates or any number of other use cases,
           | yes, the M is critical. When comparing architectures (or when
           | talking about a single architecture with the current implicit
           | conversation happening that whatever you say is meant to be
           | taken with respect to other architectures) it's conventional
           | to note that the M term always appears as a multiplicative
           | factor beyond the rest of the costs. You drop it because it's
           | obviously there but it's just visual noise that doesn't help
           | you understand where the models sit with respect to each
           | other. You only include it if it's relevant, like if you
           | found a way to get similar performance with a smaller M.
           | 
           | It's kind of like how we always drop the log term on a hash
           | table (equivalently, we work in the "word RAM" model of
           | computation), but bignum algorithm papers carefully consider
           | every log(log(log(N))) that crops up. It's relevant in some
           | contexts and not in others. You can't ever include enough
           | context to fully represent your thoughts to another person,
           | so you truncate somewhere, hopefully capturing the essence of
           | what matters.
        
       | bbor wrote:
       | The Google CEO said that this (unexpected!) invention is more
       | important than fire or electricity. In that light, I think it's
       | arrogant to try to soothsay like this. Which is tough advice
       | cause this is otherwise great analysis, and acknowledging the
       | great unknown feels scary and of dubious value. But I think it's
       | the only honest way forward
        
       ___________________________________________________________________
       (page generated 2024-02-21 23:01 UTC)