Why do you think OpenAPI is so far out in front? It's not like there's a lot of secret sauce here - most of this stuff (transformers,etc.) is all out there in papers. And places like Google & Meta must have a lot more computing resources to train on that OpenAI does thus they should be able to train faster. Do you think OpenAI has discovered something they haven't been open about?
> Why do you think OpenAPI is so far out in front?
There is a network effect forming around its models. The strengths of its kit speak for themselves. (It also cannot be understated how making ChatGPT public, something its competitors were too feeble, incompetent and behind the curve to do, dealt OpenAI a massive first-mover advantage.)
But as others note, other models are in the ballpark. Where OpenAI is different is in the ecosystem of marketing literature, contracts, code and e.g. prompt engineers being written and trained with GPT in mind. That introduces a subtle switching cost, and not-so-subtle platform advantage, that–barring a Google-scale bout of incompetence–OpenAI is set to retain for some time.
Doubtful "feeble and incompetent" are fitting considering much of the research behind OpenAI was pioneered and published by those predecessors. More like "unwilling", and perhaps for good reason. Time will tell.
The level of denial people are willing to sink into regarding how good GPT-4 is compared to everything else is truly crazy. Not a single other project is an order of magnitude close to the quantitative and qualitative (actual experiential results, not just benchmarks) results that GPT-4 brings.
I feel that there’s significant insecurity among a lot of coders about GPT-4. A lot of them are ignoring the pace of improvement and highlighting the few off chances where it gets things wrong.
I think there's a lot of people writing boilerplate programs who are going to be freed from these menial tasks (i.e. no more Java enterprise application development, thankfully).
GPT 4 is quite astounding. It might be wrong on occasion, but it will easily point you in the right direction most of the time. It still messes up, but like a twentieth of what 3.5 did. Honestly it is like an incredible rubber ducky for me. Not only can I just talk like I’m talking to a rubber duck but I can get fast, mostly informed, feedback that unblocks me. If I have a bunch of things competing for my attention I can ask gpt about one of them, a hard one, go do something else while it types out its answer, and then come back later and move on with that project.
I'm a recent convert, been experimenting with converting one PL to another.
GPT-3.5 will get the gist of what the code is doing, and then provide what looks like a direct translation but differs in numerous details whilst having a bunch of other problems.
GPT-4 does a correct translation, almost every time.
It kills me that there's a waiting list for the API. I have put together some tools to integrate 3.5 into my workflow and it helps for my current task a lot (for others it's useless). But to really shine it needs to have API access to 4.
I recently finally got access to 4 in the API, it's good. It's much better imo at following the system prompt too. Faster than you see in chatgpt I think, not as fast as 3.5-turbo but definitely less tedious.
My only kind of quantitative answer is that I had 3.5 creating ember templates & data and it would get the templates mostly ok after a couple of iterations of fixing errors, almost never on the first shot (if ever). Often wouldn't quite get it in two, and data structures would often be kinda there but not quite. Required a lot more care with the prompts. 4 gave me working things every time first time (except only where it did things that were fine in a template but not currently supported by the custom framework I'm using them in), and didn't need as much hand holding.
Qualitatively, it's wildly different from gpt-3.5-turbo for discussions. 3.5 feels a little formulaic after a while with some kinds of questions. 4 is much more like talking to an intelligent person. It's not perfect, but I'm flipping between discussing a sporting thing, then medical malpractice, legal issues, technical specifications and it's doing extremely well.
If it's affordable for you, I'd really recommend trying it.
For code 90%+ reduction is easily correct. For text and other content, I can't say, but I would guess it's not that good.
Anything involving reasoning, code, complex logic, GPT-4 is a breakthrough. GPT-3.5 turbo is more than good enough for poetry and the other text generation stuff.
My favorite part about GPT-4 is that if it generates code that is wrong, and you ask it to verify what it just wrote - without telling it whether it's wrong or not, much less pointing out the specific issue - more often than not it will spot the problem and fix it right away.
And yes, it does indeed make an amazing rubber duck for brainstorming.
GPT-4 is leaps ahead, and it's improving with every new release. The latest March 23 release is significantly better than the previous one and does a LOT of heavy lifting for my code at least.
At the very least, it's a massive productivity booster.
I've had decent success with Open Assistant, an open source model. I'd say it's within the order of magnitude of ChatGPT, given the prompts I'm looking at, including reasoning prompts. This, I believe, is due to the overwhelmingly clean data that OA have managed to acquire through human volunteers.
I dont work at Google, I think other FAANGS underinvested in this area as they didnt think it was promising. But I will admit, I am suspicious that Google is incompetent. Probably they can come back given how much money they will be forced to throw at it. But Bard is clearly behind and I dont believe their "abundance of caution" arguments for why Bard cant even code.
Google was built by brilliant people whose mission was to make any information available to anyone, instantly.
Google is run by smart people whose mission is to maximize clicks on ads. If a user finds what they’re looking for quickly, that’s lost revenue.
Google’s profit motives are not aligned with useful AI. The better AI is, the less people need to click through to lots of web pages and ads, the less revenue for Google.
I don’t think they can catch up without a major pivot in business model. It’s very hard to be deeply invested in providing more value if it means reducing your revenue.
FAANG companies have distribution. They can sell anything. Give them a bit of time. They are not companies of a single product, they milk people better than governments.
The impression I get is that they purposely limited Bard in certain ways. It does not seem to be willing to code and makes a lot of excuses when you ask it to (at least as of a couple of weeks ago when I last tried it). Did they put this limitation on it because it's not so good at coding or because they don't want it to be abused in some way? In my experience I had to trick it into coding by telling it to pretend that it had created it's own programming language and then implement an algorithm in it. It seemed to think that it had created Python.
> I am suspicious that Google is incompetent.
Google has put a lot of effort and investment into AI. With Bard I get the feeling they're not showing us what they really have - it's like for some reason they're holding back the good stuff, at least that's my suspicion.
I am still wondering if Google is experiencing it Kodak moment.
They have the dominant product that makes them billions and billions of dollars at 'relatively' low cost.
The new dominant product is on its way, but it costs far more to operate and will net them far less money, so... um no one wants to kill the goose that is still laying golden eggs, even though its days are numbered already.
(moment worth photographing): From an Eastman Kodak Company advertising campaign.
(business's failure to foresee): In reference to the Eastman Kodak Company's decline when cameras and film were overtaken by smartphones and digital technologies.
Noun
Kodak moment (plural Kodak moments)
(informal) A sentimental or charming moment worthy of capturing in a photograph.
(informal) The situation in which a business fails to foresee changes within its industry and drops from a market-dominant position to being a minor player or declares bankruptcy.
Kodak Film Commercial - These are the Moments - Baby (1993):
I don't see ads as a big problem with ChatGPT. You could put a side-bar on the right and on the fly recommend products relevant to the on-going conversation.
The cost of computing these ads would be a lot more than today's keyword-based approach, that's certainly a problem. But think of hyper-relevant ads, based on the chat itself. There's a lot of information there, that beats tracking people's behavior online all day.
That depends on ad publishers, right? If they want to sell A, B and C and I am interested in D, then Google's still showing one of A, B or C to me. D doesn't make profit if there is nobody paying for ads.
Google is advertising things we don't need, that's why ad clicks are so abysmal. LLMs won't change that.
I’ve been infuriated with DuckDuckGo on occasion because it refused to exclude certain results.
In fact when you add an exclusion clause it simply boosts those results further instead of removing them.
I’ve been told this is because the underlying search providers refuse to exclude paying customer even when you explicitly don’t want to hear from them.
I could definitely see this happening in LLM answers too and I don’t expect it to be particularly subtle.
100% this -- a Google search requires multiple input queries, refinement, and scrolling a list of possible answers that are really just links to other web sites. The ChatGPT experience is far superior to this, for the average consumer and getting close for the power user. It's a better way to ask the Internet what it knows with a more natural interface that everyone already knows how to use -- real natural language. Less cognitive overhead, no busy search results that require clicking back and forth and (for now) no ads. That last part is key -- ChatGPT is doing Google's job right now, and not even having to run ads. Google doesn't even offer a premium no ads option for search and if they did I doubt enough people would buy it anyway to matter.
If I was Google I'd be worried. Very worried indeed. They either need to dramatically change their entire company within 18 months, or accept they are going to loose substantial amount of market -- and once its gone, it's gone in a first mover, winner takes all environment like what we have right now. Just ask Google themselves what it felt like back in the early 2000's when they completely destroyed the other search engines.
Now it is. Google used to be good too, until ads started looking like search results, and then the first page became entirely ads.
In the future, when you ask ChatGPT to help you write your resume, it will try to upsell you a premium account in linked in. It will withhold its best resume advice only for LinkedIn premium users after all.
> With Bard I get the feeling they're not showing us what they really have
I highly doubt this. If they had it they would show it because if they don't react swiftly and decisively their brand will be in 'catch up' mode rather than out front where they are used to being.
It's not just about hardware, it's about the software infrastructure to go with it. Other than OpenAI most researchers weren't interested in "merely" scaling things up because that was/is seen as simple engineering, unworthy of the great minds who dream up new algorithms.
I don't know about GPT4, but GPT3.5 I'd bet is pretty traditional and boring. It's power comes from a really good, properly curated dataset (including the RLHF).
GPT3.5 turbo is much more interesting probably, because they seem to have found out how to make it much more efficient (some kind of distillation?).
GPT4 if I had to make a very rough guess, probably flash attention, 100% of the (useful) internet/books for it's dataset, and highly optimized hyperparameters.
I'd say with GPT4 they probably reached the limit of how big the dataset can be, because they are already using all the data that exists. Thus for GPT5 they'll have to scale in other ways.
To be fair, if the opposite were true, it might not be wise to admit. Saturating available high quality training data is one of the few ways anyone can see OpenAI slowing down.
1. They would already be using everything they can get
2. They would easily be able to explain what they're not using, without giving away sensitive secrets.
I wonder if we saw the same video - or maybe it is just ChatGPT being "great" in the wild? I see one guy asking another guy simple questions and getting weaselwords for an answer.
Right, that's totally it. I came away thinking the interviewer seemed way sharper than the interviewee which is pretty rare. The sheer throughput and speed of interesting questions was incredible. Too bad that many of the answers were not.
> I'd say with GPT4 they probably reached the limit of how big the dataset can be
I’m curious about this too; not just on the dataset size, but also the model size. My hunch is that the rapid improvements of the underlying model by making it bigger/giving it more data will slow, and there’ll be more focus on shrinking the models/other optimisations.
I don't think we're anywhere close to the limit of sheer hardware scalability on this. Returns are diminishing, but if GPT-4 (with its 8+ k context window) is any indication, even those diminishing returns are still very worthwhile.
If anything, I wonder if the actual limit that'll be hit first will be the global manufacturing capacity for relevant hardware. Check out the stock price of NVDA since last October.
According to financial reports they are building a $225 million supercomputer for AI. What we can probably expect is the same dataset with even more compute ran on it.
There is a soft limit due to the computation required; the currently used model architectures are quadratic with respect to context size, so if you want ten times larger context size, that's going to need a hundred times more effort.
For a little more than a year I worked in an AI startup doing basically everything other than AI (APIs, webapps, devops...), but from what I've seen there the "secret sauce" to AI success is the training process (dataset, parameters, fine-tuning steps, ...). And OpenAI isn't open about theirs since their beginnings.
> Do you think OpenAI has discovered something they haven't been open about?
They have not, which makes me curious about which company gp works for because the "F" and "G" in FAANG are publicly known to already have LLMs. Not sure about Amazon, but I'm guessing they do too.
As an outsider, the amazing thing about ML/AI research is that you get a revolutionary discovery of a technique or refinement that changes everything, and a few months later another seminal paper is published[0]. My bet is ChatGPT is not the last word in AI, and OpenAI will not have a monopoly on upcoming discoveries that will improve the state of the art. They will have to contend with the fact that Google, Meta & Amazon own their datacenters and can likely train models for cheaper[1] than what Microsoft is paying itself via their investment in OpenAI.
0. In no particular order: Deep learning, GANs, Transformers, transfer learning, Style Transfer, auto-encoders, BERT, LLMs. Betting the farm on LLMs doesn't sound like a reasonable thing to do - not saying that's what OpenAI is doing, but there are a lot of folk on HN who are treating LLMs as the holy grail.
1. OpenAI may get a discount, but my prediction when they burn through Microsoft, they'll end up being "owned" by Microsoft for all intents and purposes.
Issue is the data moat OAI is building. They'll have hundreds of millions of high quality user interactions with ChatGPT they can use to finetune their models. What will anyone else including Google have?
Google has been collecting user interactions since 2007 via GOOG-411, which was a precursor to the Google Assistant - I suspect Google has billions of user interactions on hand through the latter. Facebook has posts and comment, Amazon has products pages, reviews and product Q&As and all of them have billions of dollars to draw upon if they choose to buy high-quality data, or spin-up / increase teams that create and/or categorize training data.
They also have deep roster of AI researchers[1] to potentially obsolete LLMs or make fine-tuning work without access to of ChatGPT records.
1. I suspect Google alone has more AI researchers that OpenAI has employees
I'm not sure how deep that moat is. As soon as you open up the API, anyone can distil ChatGPT (or at least, some smaller part of it) by fine-tuning another model on its outputs[0].
I'm guessing that this is the #1 fear for people inside OpenAI have right now.
[0] For the record, I have zero problem with this.
1. ToS make it hard for a commercial entity to do so. So some third parties would have to collect the data first
2. You won't be able to get the hundreds of millions or more interactions that OAI will have (both due to cost of API as well as it being not easy to figure out a good way to generate that many queries for a good multiturn conversaton). Maybe you can make up for it by querying smartly. We don't know if we can right now.
As people make chat bots with openAI and tie them into existing chat services, organizations that offer these chat services will get their hands on that kind of data too.
A lot of FAANG data folks aren't on the teams that were doing research into this stuff and weren't using the latest fruits of that research.
OpenAI has released a ton more easy-to-use-for-everyone stuff that has really leapfrogged what a lot of "applied" folks everywhere else were trying to build themselves, despite being on-the-face-of-it more "general."
I think it’s the way things go usually. The big players have a business to run so they can’t focus much on innovation. OpenAI has the only purpose right now to push AI and nothing else. Once they have a real business they will also slow down.
They have been collecting human feedback data for 2 years + probably have a lot of data from Copilot + are training with large context models + have invested a ridiculous amount in curating pretraining data -- the kind of stuff that won't get you a ton of pubs (so you won't see Google researchers having focused on it a lot) but apparently turns out to be super important for a good LLM
All of the neural network architecture for human level thinking and processing, including vision, speech, emotion, abstract thought, balance and fine motor skills, everything was publicly released in April 2003, twenty years ago this month. It's a 700 megabyte tarball and sets up an 80b parameter neural network.
What? Huh? Yes the human genome encodes all human level thought.[1] Clearly it does because the only difference between humans that have abstract thought as well as language capabilities and primates that don't is slightly different DNA.
In other words: those slight differences matter.
To anyone who has used GPT since ChatGPT's public release in November and who pays to use GPT 4 now, it is clear that GPT 4 is a lot smarter than 3 was.
However, to the select few who see an ocean in a drop of water, the November release already showed glimmers of abstract thought, many other people dismiss it as an illusion.
To a select few, it is apparent that OpenAI have found the magic parameters. Everything after that is just fine tuning.
Is it any surprise that without OpenAI releasing their weights, models, or training data, Google can't just come up with its own? Why should they when without turning it into weights and models, the human neural network architecture itself is still unmatched (even by OpenAI) despite being digitized twenty years ago?
No, it's no surprise. OpenAI performed what amounts to a miracle, ten years ahead of schedule, and didn't tell anyone how they did it.
If you work for another company, such as Google, don't be surprised that you are ten years behind. After all, the magic formula had been gathering dust on a CD-ROM for 20 years (human DNA which encodes the human neural network architecture), and nobody made the slightest tangible progress toward it until OpenAI brute forced a solution using $1 billion of Azure GPU's that Microsoft poured into OpenAI in 2019.
Is your team using $1 billion of GPU's for 3 years? If not, don't expect to catch up with OpenAI's November miracle.
p.s. two months after the November miracle, Microsoft closed a $10 billion follow-on investment in OpenAI.
Until we see otherwise, don't we have to assume there's some secret sauce? Bard doesn't match GPT4 and it isn't for a lack of trying. (though perhaps that will change, so far that's the case)
Bard would not trick anyone into thinking it was sentient, yet something they have supposedly did. I just think Google has far more to lose than Bing, so they are being more cautious.
> Bard would not trick anyone into thinking it was sentient, yet something they have supposedly did
Bard is overtly a reduced-resources model compared to the best version of the same technology (which, if true, is probably a boneheadedly bad choice for a public demo when everyone is already wowed by the people who got theirs out first, but easily explains that disparity. Though so does “guy who wanted public attention made stuff up well-calibrated to that goal.”)
> Bard is overtly a reduced-resources model compared to the best version of the same technology
There's a scaling problem. ChatGPT/LLM systems cost far more to run per query than the Google search engine. Google can't afford to make those the first line query reply.
A big business model question is whether Google will insist you be logged in to get to the large language model.
At Google scale, these things are going to have to be a hierarchy. Not everything needs to go to a full LLM system. Most Google queries by volume can be answered from a a cache.
> Most Google queries by volume can be answered from a a cache.
And given how aggressively they limit the number of search results (in spite of listing some ridiculous number of results on page #1) that percentage may well be very large.
Google also has Sundar and Ruth who’d rather focus on how to get another ad on the SERP than kill the golden goose. They’re not going slow, they just don’t have the leadership for the moment.
Maybe it is just time to train and also good training data for prompts which openai has gathered for so long already? E.g. there is a bottleneck on how fast you can train and also gather good data.
Possibly, but wouldn't Google and Meta have access to way more compute resources and data than OpenAI? Google has been touting their TPUs for several years now.
OpenAI has access to Microsoft and Azure. That’s bigger than Meta, roughly on par with Google in terms of capability and higher in terms of market cap.
Google has the compute, from the comparisons I have seen Bard smokes GPT-3.5-Turbo on response times. So my guess is that internal politics prevents them from putting out something better. There would have to be immense pressure from the search division to not make them obsolete.
Without Nadella footing the compute bills, nobody would be taking about OpenAI. He’s brilliant, he let the start up take on huge risk to quietly claim the gains for m$.
If you try Bard or Claude or character.ai they are not far behind GPT4. They might even be on par in terms of raw LLM capabilities. ChatGPT has better marketing and in some cases better UX. A lot of this is self-fulfilling. We think it's far ahead, so it appears to be far ahead.
> If you try Bard or Claude or character.ai they are not far behind GPT4
Bard is way behind ChatGPT with GPT-3.5, much less GPT-4. Haven’t tried the others, though.
OTOH, that’s way behind qualitatively, not in terms of time-of-progress. So I don’t think it is at all an insurmountable lead, as much as it is a big utility gap.
Claude and Character AI are great at holding a conversation but they lack the ability to do anything specialized that really makes these LLM’s useful in my day to day life. I ask GPT-4 and ChatGPT questions I would ask in stackoverflow, I can’t do that with Claude or Character AI. Bard actually seems behind even conversationally to the rest
