The real sauce is the data you fine tune these foundation models on, so RLHF, specific proprietary data for your subfield, etc. The model definition, basically Transformer architecture and a bunch of tricks to get it to scale are mostly all published material, hyper parameters to train the model are less accessible but also part of published literature; then the data and (probably) niche field you apply it to becomes the key. Gonna be fun times!

For example, the biggest LLama model only runs on an A100 that costs about $15,000 on ebay. The new H100 that is 3x faster goes for about $40,000 and both of these cards can only support a limited number of users, not the tens of thousands of users who can run off a high-end webserver.

I'd imagine Google would lose a lot of money if they put GPT4 level AI into every search, and they are obsessed with cost per search. Multiply that by the billions and it's the kind of thing that will not be cheap enough to be ad supported.

Sign me up for that kickstarter!

It sounds like there is a lot of work happening on sparse networks now, so it’ll be interesting to see how this changes in the near future.

I'm not saying it's impossible but if resources allow it makes a lot of sense to start with the biggest model you can still train. Especially since for whatever reason things seem to get a lot easier if you simply throw more computing power at it (kind of like how no matter how advanced your caching algorithm it's not going to be more than 2 times faster than the simplest LRU algorithm with double the amount of cache).

There has been a lot of work on sparsity and discovering sparse subnetworks in trained dense networks. And intel even proposed some alternative cpu friendly architectures and torch/tf and gpus are starting to do okay with sparse matrixes so thing are changing.

Google, in collaboration with OpenAI, has published an impressive tour de force where they have throughly developed and validated at scale a sparse transformer architecture, applied to general language modeling task: https://arxiv.org/abs/2111.12763

This happened in November of 2021, and there is a public implementation of this architecture on the Google's public github.

Impressively, due to some reasons, other up-and-coming players are still not releasing models trained with this approach, even though it promises multiplicative payoff in inference economy. One boring explanation is conservatism for NN training at scale, where training runs cost O(yearly salary).

Let's hope the open source side of things catches up.

Of course from a memory bandwidth and model size perspective maybe there are benefits long before that.

Benchmarks: https://github.com/ggerganov/llama.cpp/issues/34

A 4-bit quantize of a 1000B model should be <600GB, so would fit on a regular 8x80 DGX system.

https://github.com/ggerganov/llama.cpp/issues/34

An M1 Max does 100ms per token. A 64 core threadripper about 33ms per token.

If I want a response within 100ms, and have 1000 AI-queries per day, that would only be about 2 minutes of aggregated processing time for your AI box per day. It's less than 1% utilised. If the same box is multiuser and on the internet, it can probably serve 50-100 peoples queries concurrently.

The converse is that if you put something onto the cloud, for the same cost you might be able to effectively get 50x the hardware per user for the same cost (i.e. rather than have 1 AI box locally with 1 GPU for each of the 50 users, you could have 1 AI box with 50 GPU's which is usable by all 50 users).

Maybe this is why Sam Altman talked about "the end of the large LLMs is here"? He understands anything bigger than ChatGPT-4 isn't viable to run at scale and be profitable?

At what limit does it stop getting better?

At the time, operating YouTube was eye wateringly expensive and lost billions. But google could see where things were going: a triple trend of falling storage costs, falling bandwidth and transmission costs (I’m trying to dig up a link I read years ago about this but google search has gotten so shit that I can’t find it).

It was similar for Asic miners for Bitcoin. Given enough demand, specialised, lower cost hardware specially for LLMs will emerge.

[0] https://cppcast.com/ai_infrastructure/ [1] https://github.com/orgs/unum-cloud/repositories?type=all

The 3060 effectively matched a 1080 at $329 in 2021 (and has 50% more VRAM at 12GB instead of the 1080's 8GB), so call it five years if the 3070 isn't mid-range enough.

The 3060 Ti launched in 2022 at $399 and handily beat the 1080 on benchmarks, so call it six years if you want the midrange card to beat (not just match) the previous top-of-the-line card, and if a *70 card doesn't count as midrange enough. Less than a decade still seems like a reasonable claim for a midrange card to beat a top-of-the-line card.

Taking 6 years to double the perf/$ implies that it would take ~42 years for a $40000 H100 to reach mid-range levels. Assuming scaling, particularly VRAM, holds.

And plus it would be getting really close to the Landauer limit by that point.

(1) make a whole lot of money fulfilling that demand, or

(2) leave an unmet demand that makes it more attractive for someone else to spend the money to field a solution.

(1) seems an attractive choice even if it wasn’t for the potential threat of (2).

That's a big if. I imagine people who will buy GPUs to run LLMs locally are either researchers or programmers. I am imagine most consumer focused solutions will be cloud first. That is a much smaller market than gamers and nvidia wouldn't want to cannibalize their datacenter offerings by releasing something cheaper. It's far better for them to sell high tier GPUs to Amazon and let them "rent" it out to researchers and programmers.

client side tokens per second will be through the roof and the models will be smaller

-EQ moment

edit: its about the combination of those methods making models accessible on consumer hardware

Or at about a 10x speed reduction you can run it on 128 GB of RAM for only around $250.

The story is not anywhere near as bleak as you paint.

https://github.com/ggerganov/llama.cpp/issues/34

You need to load the data so the graphics cards - where the compute is - can use it to answer queries. But you don’t need a separate copy of the data for each GPU core, and though slower, cards can share RAM. And yet even with parallel cores, your server can only answer or process so many queries at a time before it runs out of compute resources. Each query isn’t instant either given how the GPT4 answers stream in real-time yet still take a minute or so. Plus the way the cores work, it likely takes more than one core to answer a given question, likely hundreds of cores computing probabilities in parallel or something.

I don’t actually know any of the details myself, but I did do some CUDA programming back in the day. The expensive part is often because the GPU doesn’t share memory with the CPU, and to get any value at all from the GPU to process data at speed you have to transfer all the data to GPU RAM before doing anything with the GPU cores…

Things probably change quite a bit with a system on a chip design, where memory and CPU/GPU cores are closer, of course. The slow part for basic replacement of CPU with GPU always seemed to be transferring data to the GPU, hence why some have suggested the GPU be embedded directly on the motherboard, replacing it, and just put the CPU and USB on the graphics card directly.

Come to think of it, an easier answer is how much work can you do in parallel on your laptop before you need another computer to scale the workload? It’s probably like that. It’s likely that requests take different amounts of computation - some words might be easier to compute than others, maybe data is local and faster to access or the probability is 100% or something. I bet it’s been easier to use the cloud to toss more machines at the problem than to work on how it might scale more efficiently too.

The theoretical best approach would be to integrate lots of GPU cores on the motherboard alongside very fast memory/storage combos such as Octane, but reality is very different because we also want portable, replaceable parts and need to worry about silly things like cooling trade offs between placing things closer for data efficiency vs keeping things spaced apart enough so the metal doesn’t melt from the power demands in such a small space. And whenever someone says “this is the best graphics card,” someone inevitably comes up with a newer arrangement of transistors that is even faster.

The factor is basically 2 * number of layers * number of embeddings values (e.g. fp16) that are stored per token.

I still think there is a lot to be gained from just properly and efficiently composing the parts we already have (like how the community handled stable diffusion) and exposing them in an accessible manner. I think that’ll take years even if the low hanging algorithm fruits start thinning out.

...that's probably Koomey's law[1][3], which looks well on track to hold for the rest of our careers. But eventually as computing approaches the Landauer limit[2] it must asymptotically level off as well. Probably starting around year 2050. Then we'll need to actually start "doing more with less" and minimizing the number of computations done for specific tasks. That will begin a very very productive time for custom silicon that is very task-specialized and low-level algorithmic optimization.

[0] Shows that Moore's law (green line) is expected to start leveling off soon, but it has not yet slowed down. It also shows Koomey's law (orange line) holding indefinitely. Fun fact, if Koomey's law holds, we'll have exaflop power in <20W in about 20 years. That's equivalent to a whole OpenAI/DeepMind-worth of power in every smartphone.

The neural engine in the A16 bionic on the latest iPhones can perform 17 TOPS. The A100 is about 1250 TOPS. Both these performance metrics are very subject to how you measure them, and I'm absolutely not sure I'm comparing apples to bananas properly. However, we'd expect the iPhone has reached its maximum thermal load. So without increasing power use, it should match the A100 in about 6 to 7 doublings, which would be about 11 years. In 20 years the iPhone would be expected to reach the performance of approximately 1000 A100's.

At which point anyone will be able to train a GPT-4 in their pocket in a matter of days.

There's some argument to be made that Koomey himself declared in 2016 that his law was dead[4], but that was during a particularly "slump-y" era of semiconductor manufacturing. IMHO, the 2016 analysis misses the A11 Bionic through A16 Bionic and M1 and M2 processors -- which instantly blew way past their competitors, breaking the temporary slump around 2016 and reverting us back to the mean slope. Mainly note that now they're analyzing only "supercomputers" and honestly that arena has changed, where quite a bit of the HPC work has moved to the cloud [e.g. Graviton] (not all of it, but a lot), and I don't think they're analyzing TPU pods, which also probably have far better TOPS/watt than traditional supercomputers like the ones on top500.org.

0: (Slide 13) https://www.sec.gov/Archives/edgar/data/937966/0001193125212...

1: "The constant rate of doubling of the number of computations per joule of energy dissipated" https://en.wikipedia.org/wiki/Koomey%27s_law

2: "The thermodynamic limit for the minimum amount of energy theoretically necessary to perform an irreversible single-bit operation." https://en.wikipedia.org/wiki/Landauer%27s_principle

3: https://www.koomey.com/post/14466436072

4: https://www.koomey.com/post/153838038643

by mid 1980s personal computers costed less than $500

While it may have been true that it was technically possible to assemble a PC for $500... good luck. In the real world people were spending $1500-$2500+ for PCs, and that price point held remarkably constant. By the time you were done buying a monitor, external drives, printer etc $3000+ was likely.

https://en.m.wikipedia.org/wiki/IBM_Personal_Computer#:~:tex....

Or see Apple Mac 512 introduced at approx $2800? One reason this was interesting (if you could afford it) was the physical integration and elimination of "PC" cable spaghetti.

https://en.m.wikipedia.org/wiki/Macintosh_512K

But again having worked my way thru college with an 8 MB external hard drive... which was a huge improvement over having to swap and preload floppies in twin external disk drives just to play stupid (awesome) early video games, all of this stuff cost a lot more than youre saying. And continued to well into the 90s.

Of course there are examples of computers that cost less. I got a TI-99/4a for Christmas which cost my uncle about $500-600. But then you needed a TV to hook it up to, and a bunch of tapes too. And unless you were a nerd and wanted to program, it didn't really DO anything. I spent months manually recreating arcade video games for myself on that. Good times. Conversely, if you bought an IBM or Apple computer, odds were you were also going to spend another $1000 or more buying shrinkwrap software to run on it. Rather than writing your own.

Source: I remember.

The CPC 464 is the first personal home computer built by Amstrad in 1984. It was one of the bestselling and best produced microcomputers, with more than 2 million units sold in Europe.

Price

£199 (with green monitor), £299 (with colour monitor)

> But again having worked my way thru college with an 8 MB external hard drive

that was a minicomputer at the time, not a PC (personal computer)

> Source: I remember.

Source: I owned, used and programmed PCs in the 80s

I believe this was the PC of the ordinary person (In the "personal computer" sense of the word.)

It's a lower cost of entry than almost any other industry I can think of. A cargo van with a logo on it (for a delivery business or painting business, for example) would easily cost 10-20x as much.

2. The problem is scaling. To support billions of search queries you would have to invest in a lot more than a single GPU. You also wouldn't only need a single van, but once you take scaling into account even at $3500 the GPUs will be much more expensive.

That said, costs will come down eventually. The question in my mind is whether OpenAI (who already has the hardware resources and backed by Microsoft funding to boot) will be able to dominate the market to the extent that Google can't make a comeback by the time they're able to scale.

2 - Scaling is not a problem in other industries? If you want to scale your food truck, you will need more food trucks, this doesn't seem to really do anything for your point.

GGML and GPTQ have already revolutionised the situation, and now there are tiny models with insane quality as well, that can run on a conventional CPU.

I don't think you have any idea what is happening around you, and this is not me being nasty, just go and take a look at how exponential this development is and you will realise that you need to get in on it before its too late.

Even so, I am able to produce insane results locally with open source efforts on my RTX3060, and now I am starting to feel confident enough that I could take this to the next level by either using cloud (computerender.com for images) or something like vast.ai to run my inference (or even training if I spend more time learning). And if that goes well I will feel confident going to the next step, which is getting an actual SOTA GPU. But that will only happen once I have gained sufficient confidence that the investment will be worthwhile. Regardless, apologies for suggesting the RTX3060 is SOTA, but to me in a 3rd World Country, being able to run vicuna13b entirely on my 3060 with reasonable inference rates is revolutionary.

GP lives in an company world. The cost of a developer to a company is the developer's salary as stated in the contract, plus some taxes, health insurance, pension, whatever, plus the office rent for the developer's desk/office, plus the hardware used, plus a fraction of the cost of HR staff and offices, cleaning staff, lunch staff... it adds up. $3500 isn't a lot for a week.

So you get situations where someone names a number and someone else reacts by thinking it's horribly, unrealistically high: The former person thinks in employer terms, the latter in employee terms.

I live in the real world, at a small company with <100 employees, a thousand miles away from SV.

$3200 * 52 == $180k a year, and gives $120k salary and $60k for taxes, social security, insurance, and other benefits, which isn't nearly FAANG level.

Even if you cut it in half and say it's 2 weeks of dev salary, or 3 weeks after taxes, it's not unreasonable as a business expense. It's less than a single license for some CAD software.

> 2. The problem is scaling. To support billions of search queries you would have to invest in a lot more than a single GPU. You also wouldn't only need a single van, but once you take scaling into account even at $3500 the GPUs will be much more expensive.

Sure, but you don't start out with a fleet of vans, and you wouldn't start out with a "fleet" of GPUs. A smart business would start small and use their income to grow.

Which puts it out of reach for the majority of people, and means many people who could technically afford it will choose not to spend that much on it anyway.

It's only a matter of time before someone makes a similar device with a decent LLM on it, and premium ones will have more memory/cpu power.

I know exactly what you're talking about because my father-in-law had the same thing. I'm just very skeptical that specialist hardware will overtake general commoditized computing devices for mass-market usage. The economics alone make it unlikely.

One of them has an unlimited "free forever" internet subscription so it can fetch network translations online if the local dictionary doesn't have the word.

[0] even other tech developers who, like me, migrated somewhere where a language barrier came up

Wiki says: https://en.wikipedia.org/wiki/Ot%C3%A1vio_Good

    To develop Word Lens, Otávio Good founded Quest Visual Inc., which was acquired by Google, Inc. in 2014, leading to the incorporation of the Word Lens feature into the Google Translate app in 2015.

The people I show it to are surprised that it's possible in 2023, even though it was demoed at the end of 2010.

By comparison, the Cray is outperformed by my smartphone.

Actually, it is outperformed by my previous smartphone, which I purchased in 2016 and replaced in 2018.

Actually, it is outperformed by a single core on my previous smartphone, of which it has eight cores.

Besides what Google does here is besides the point, because Bing has already unleashed ai search. Google will either follow along or stop being relevant.

I think OP is arguing that in that partnership Microsoft holds the power, as they have the existing platforms. The linked article argues that AI technology itself is not as much of a moat as previously thought, and the argument therefore is that Microsoft likely doesn't need OpenAI in the long term.

It seems incredibly likely that the FTC will block that. New leadership seems to be of the opinion that consumer harm is the wrong standard. Buying the competition with profits from a search monopoly leaves all parties impoverished.

Anyways, I don't think the risk is failure, but of non-success. The article claims meta won but it seems like nvidia is the winner: everyone uses their chipsets for training, fine tuning and inference. And the more entrants and niche applications show up the more demand there is for their product. TPUs theoretically play into this, but the "leak" doesn't mention them at all.

Like the saying goes: during a gold rush, sell shovels.

There’s no path forward for Google which doesn’t involve firing a lot of managers and replacing them with people who think their income depends on being a lot better at building and especially maintaining products.

Yes, but the quickest way for anyone to get themselves to state-of-the-art is to buy OpenAI. Their existential risk is whether they continue to be (semi)independent, not whether they shutdown or not. Presumably Microsoft is the obvious acquirer, but there must be a bunch of others who could also be in the running.

https://www.semianalysis.com/p/google-we-have-no-moat-and-ne...

I...kinda don't want this? UIs have already changed in so many different fits, starts, waves, and cycles. I used to have skills. But I have no skills now. Nothing works like it used to. Yeah they were tricky to use but I cannot imagine that a murky AI interface is going to be any easier to use, and certainly impossible to master.

Even if it is easier to use, I am not sure I want that either. I don't know where the buttons are. I don't know what I can do and what I can't. And it won't stay the same, dodging my feckless attempts to commit to memory how it works and get better at it...?

A lot of people hate it as it’s closer to Google assistant than GPT4 and it makes mechanical noise when it rotates but I don’t think it’s a terrible idea.

Anthropomorphism, the attribution of human traits to things, is common with cars.

I would like my car to run a LLM instead of being so stupid at understanding my voice commands.

https://twitter.com/jasminezroberts/status/16056114516740259...

With modern software, I find it that there isn't much to learn at all - in the past decade, seems to only be removing features and interaction modes, never adding anything new.

Still, I don't regret having learned so much all those years ago. It gives me an idea what the software could do. What it was supposed to do. This means I often think of multi-step solutions for problems most people around me can't solve unless there's a dedicated SaaS for it. As frustrating as it often is to not be able to do something you could 10 years ago, sometimes I discover that some of the more advanced features still remain in modern toy-like software.

"In 2003, I was given the chance to lead the redesign of the most well-known suite of productivity software in the world: Microsoft Office.

"Every day, over a billion people depended on apps like Word, Excel, PowerPoint, and Outlook, so it was a daunting task. This was the first redesign in the history of Office, and the work that we did ended up shaping the standard productivity experience for the next two decades."

...

https://jensenharris.com/home/office

https://jensenharris.com/home/ribbon

I’m still waiting for people to stop complaining about it.

Originally: Taxonomically organized nested menus, culminating at a specific function or option screen

Now: Usage-optimized Ribbon (aka Huffman coding for the set off all options), culminating at a specific function or option screen

Future: LLM powered semantic search across all options and actions, generating the exact change you want to implement

Why have an "email signature" options page at all, when an LLM can stitch together the calls required to change it, invoked directly from English text?

10-20 years from now? Maybe. It depends on whether or not the industry will cut corners in this part of the experience. I find it hard to predict which features get polished, and which are forever left barely-functioning. Might depend on which are on "critical path" to selling subscriptions.

0-10 years from now? We'll still need the options page.

"Email signature" options page provides visibility, consistency, and predictability. That is, you can see what your e-mail signatures are, you are certain those are the ones that will be used for your message, under conditions defined in the UI pane, and if you change something, you know exactly how it will affect your future e-mails.

LLMs are quite good at handling input, though not yet reliable enough. However, as GUI replacement, they are ill-suited for providing output - they would be describing what the result is, whereas the GUI of today display the result directly. As the writers' adage goes, "show, don't tell".

(That said, the adage is way overused in actual storytelling.)

Think less "make a signature for me" and more "here's the signature I want to use, make it my default."

Then the model would map to either Outlook / Options / Signature / fields or directly to whateverSetSignature().

From that more modest routing requirement, it seems a slam dunk for even current models (retrained to generate options paths / function calls rather than English text).

LLMs could certainly help loosen the input requirements, not to mention aim some sorely needed hype back in this direction. I am afraid that they will murder the latency and compute requirements, but hey, progress is always two steps forward one step back.

For a while, ubuntu had a local search where you could hit a button (super?) start typing, and it would drill through menus at lightning speed

Maybe I'll reject instead of embrace the next big thing once I'm old enough?

All the fancy dialogs will switch around every few years.

This mindset extends to stuff like Word. Whenever you do something think hard about the essence of what you’re doing. Realize this should have been a script, but due to constraints in reality you are forced to use some wanky GUI.

If you look at it like this, you won’t care the pixels move around. Your mental model will be solid and building that is 90% of the work.

in another browser tab i'm running jupyter (the renamed ipython notebook from 02011) to graph some dsp calculations with matplotlib (02003, but mostly providing the plotting functions from matlab (01984)) which i made with python (01991, but i've only been using it since 02000) and numpy (02006, but a mostly compatible reimplementation of numeric from 01995, which i've been using since 02003). in jupyter i can format my equations in latex (01984, but for equations basically the same as plain tex82 from 01982, but i've only been using them since 01999) and my text in markdown (02004, though jupyter's implementation supports many extensions). i keep the jupyter notebook in git (02005, but i've only been using it since 02009, when i switched from cvs (01986, but i've only been using it since about 01998)). the dsp stuff is largely applications of the convolution theorem (01822 or 01912) and hogenauer filters (01981).

i do most of my programming that isn't in jupyter in gnu emacs (01984) except that i prefer to do java stuff in intellij idea, which i first used in 02006

earlier this year, my wife and i wrote our wedding invitation rsvp site in html (01990, but using a lot of stuff added up to 02000) and css2 (01998) plus a few things like corner-radius (dunno, 02006?) and a little bit of js (01995, but in practical terms 02004), with the backend done in django (02005) backed by sqlite (02000, but this was sqlite3 (02004), but sqlite mostly just implements sql (01974, first publicly available in 01979, but mostly the 01992 ansi standard) which in turn mostly just implements codd's relational data model (01970) and acid transactions (01983), all of which i've been using since 01996). and of course python, emacs, and git. most of the css debugging was done with chromium's web inspector (present from the first chrome release in 02008, a clone of firebug (02006)).

for looking up these dates just now, i used google's newish structured search results, which mostly pull information from wikipedia (02001); i also used stack overflow (02008) and its related sites.

the median of the years above is 01998, with 25% and 75% quartiles of 01987 and 02004, which i calculated using r (01997, but a reimplementation of s (01976)). if we assume that each new introduction listed above displaced some existing skill, then we can vaguely handwave at a half-life of about 25 years for these churning technology skills, which to me seems like enough time for a lot of them to amortize

The killer is being able to tell a program what you want it to do, then not having to fuddle with buttons or menus at all (unless you want to tweak things).

Then she had kids and stopped working for a while and the technology moved on without her. Now she’s like any other old person, doesn’t know how to use a computer and gets flustered when eg: trying to switch from the browser back to Word. Her kids and grandkids clown on her for being hopeless with computers.

I asked her what it was she found difficult about modern computers compared to what she worked with 50 years ago. She said it’s the multitasking. The computers she had worked with just did one thing at once.

It was when sonic went 3D that it all began for me.

Wait…he’s running away from me?

I can think of a myriad of use-cases for AI that involve custom-tuning foundation models to user-specific environments. Think of an app that can detect bad dog behavior, or an app that gives you pointers on your golf swing. The moat for AI is going to be around building user-friendly tools for fine-tuning models to domain-specific applications, and getting users to spend enough time fine-tuning those tools to where the switch-cost to another tool becomes too high.

When google complains that there is no moat, they're complaining that there is no moat big enough to sustain companies as large as Google.

There’s a reason YT or TikTok recommendation is so revered

Agreed but I don’t think the products that’ll gain market share from this wave of AI will be legacy web 2 apps; rather it’ll be AI-native or first apps that are build from ground up to collect user data and fulfill user intent. Prime example is TikTok.

It's (sort of) obviated search for certain kinds of queries engineers make, but not normies.

I say sort of, because IMO it's pretty bad at spitting out accurate (or even syntactically correct) code for any nontrivial problem. I have to give it lots of corrections, and often it will just invent new code that also is broken in some way.

Normies don't care about the exact truth

The eventual winner will have a model that stays up-to-date.

Phind is getting awfully close to this point already really. Integrating new knowledge isn't a bottleneck like we know from expert systems, I think it just hasn't been a priority for research and commercial reasons, till recently.

It's possible GPT-4 will be better; I haven't been able to test it because I remain on the waitlist.

I remain skeptical.

No need to verify the ages. You will immediately find the problem.

“List the presidents of the us in the order of their ages when they were first inaugurated”

Books had editors and were expensive to publish, which imparted some automatic credibility. You might even have involved a librarian or other expert in your search. So a lot of the credibility problem was solved for you, up-front, once you got the information source.

Google changed the game. It gave you results instantly, from sources that it guessed looked reliable. But you still had to ascertain credibility yourself. And you might even look at two or three pages on the same topic, quickly.

Google has been mostly defeated now and often none of the links it suggests are any good. That trade-off seems to be done.

Here comes LLMs. Now it's transferring even more of the work of assessing credibility to the end user. But the benefit is that you can get very tailored answers to your exact query; it's basically writing a web page just for you in real time.

I think the applications that win in this new era will have to make that part of their business model. In science fiction, AIs were infallible oracles. In the real world it looks like they'll be tireless research assistants with an incredible breadth of book-learning to start from but little understanding of the real world. So you'll have a conversation as you both converge on the answer.

If anything advances in AI & search tech will do a better job at providing citations that agree & disagree with the results given. But this can be a turtles all the way down problem.

Maybe I’m just from a different time but I can recall hearing and respecting the advice, “don’t believe everything you read on the internet”.

So, given a web page (found via search, example 1 in your comment), I remember that advice and take any suggestions with a grain of salt.

And if I’m using a social platform (I’ll reference reddit since it has historically been useful [in my experience] for interacting with other humans), typically / hopefully the useful information provided by the individual will preface their suggestions with IANAL (I’m not a lawyer) or ‘in my experience’.

In both cases, the suggestion to the query is prefaced with “this might not be right, but…”.

With my limited understanding of “AI” options that are available (to date), hinting at the fact that the information provided might not be accurate is non-existent. The entire perceived value in these systems is that the information is accurate and reliable. But that isn’t 100% the case.

So, hopefully, people will continue to take any advice with a grain of salt but it’s hard to imagine they will when the entire concept behind the “AI” idea is that the information provided is accurate.

I’ll conclude with this statement, which might negate any agreement with my statement thus far and show my disdain, but I’m not a fan of this ignorant embrace of “AI”. IMHO, it (“AI”) is embracing the idea that there are simple solutions to complex problems. I don’t think, 99% of the time, there are simple solutions to complex problems and pretending like there are is only going to create more complex problems.

Maybe I’m just being ignorant and very negative about the advances in tech but I can’t help but think folks are setting things up for failure and more confidently incorrect information.

That’s not perfect but I think it’s a lot better than building things into Word will be. There’s almost no chance that people won’t trust suggestions there more than random web searches and the quality of the writing will make people more inclined to think it’s authoritative.

Consider what happened earlier this year when professor Tyler Cowen wrote an entire blog post on a fake citation. He certainly knows better but it’s so convenient to use the LLM emission rather than do more research…

https://www.thenation.com/article/culture/internet-archive-p...

"In theory, the human labeler can include all the context they know with each prompt to teach the model to use only the existing knowledge. However, this is impossible in practice." [1] Therefore causing and forcing some connections that are not all there for the LLM. Extrapolate that across various subjects and types of queries and there you go.

1:https://huyenchip.com/2023/05/02/rlhf.html

Microsoft has a stake in OpenAI and has integrated into Azure, Bing and Microsoft 365.

In my opinion Microsoft has equal or more reach/subsidy advantage than Google for AI, at least toward general consumers.

OpenAI has no moat, but at least they have first mover advantage on a cool product, and may be able to get some chumps (microsoft) to think this will translate into a lasting feature inside of office or bing.

OpenAI just beat Google to the cool chatbot demo.

Imagine for a moment, in a different universe, in a different galaxy, another planet is ostensibly a mirror image of Earth, evolving along the same trajectory. However on this hypothetical planet, anything is possible. This has resulted in some interesting differences.

The No Google License

Neither Google, its subsidiaries, business partners nor its academic collaborators may use this software. Under no circumstance may this software be directly or indirectly used to further Google's business or other objectives.

If 100s or 1000s or more people on planet X started adopting this license for their open source projects, then of course it won't stop Google from copying them or even using the code as is. But it would muddy the waters with 100s or 1000s or more potential lawsuits. Why would any company risk it.

There is nothing stopping anyone writing software for which they have no intention of charging license fees. It's done all the time these days. There is also nothing stopping anyone from prohibiting certain companies from using it, or prohibiting certain uses.

I recall in the early days of the web when "shareware" licenses often tried to distinguish commercial from non-commercial use. Commercial use would presumably incur higher fees. Non-commercial use was either free or low cost. I always wondered, "How is the author going to discover if XYZ, LLC is using his software?" (This is before telemetry was common.) The license seemed unworkable, but that did not stop me from using the software. I was never afraid that I would be mistaken for a commercial user and the author would come knocking asking me to agree to a commercial license. I doubt I was the only one bold enough to use software with licenses prohibiting commercial use.

Even a "No Microsoft License" would make Github more interesting. One could pick some random usage. Microsoft may not this software for X. Would this make MSFT's plans more complicated. Try it and see what happens. Only way to know for sure.

Instead, MSFT is currently trying to out the plaintiffs in the Doe v Github case, over MSFT's usage of other peoples' code who put their stuff on Github, and as the Court gets ready to decide the issue, it's becoming clear IMO that if these individual are named, these brave individuals will lose their jobs and be blackballed from ever working in software again.

The No Internet Advertising License

This software may not be used to create or support internet advertising services for commercial gain.

But because any particular invention about LLMs is not a specific product but an approach, it would just be re-implemented.

One could imagine patenting an approach, if it ends up being patentable, and then giving everyone but some excluded entities a grant of royalty-free use. But, unless the use if that particular approach is inevitably very obvious (which is really unlikely with ML models), you would have hard time detecting violations and especially enforcing your patent.

I don't have faith in OpenAI as a company, but I have faith in Open-Source. What you're trying to say, if I understand correctly, is that Google will absorb the open-source and simply be back on top. But who will maintain this newly acquired status quo for Google? Google cannot EEE their own developer base. They said that much in the article;

> We cannot hope to both drive innovation and control it.

History as an example, Android did not kill *nix. Chrome did not kill Firefox. Google Docs has not killed Open Office. For the simple fact that Google needs all of these organizations to push Google forward. Whether that means Google gets access to code, or whether that means Google becomes incentivized to improve in some way.

If Google wants to eat another free lunch tomorrow they have no choice but to leave some of that free labor standing, if not prop it up a little. The real question becomes, how much market share can we realistically expect without eating tomorrow's lunch?

Probably, internally Googlers are arguing that the "AI explosion" is short lived and people will be stop paying for AI as soon as open source PC models become cost and quality competitive. So they shouldn't chase the next big revenue stream that OpenAI is currently enjoying because it's short lived.

- usable by anyone

- feasible on your desktop

Thereby at least levelling the playing field for other developers.

As always, Apple's focus is hardware-first, and I think it will once again pay off here.

Nevermind more specialized tools that don't have the resources to develop their own competent AI - google might pull it off but adobe won't, and millions Saas and small programs won't even try. As another example, how could an Adobe AI ever have a better interpretation of 'Paint a pretty sunset in the style of Picasso' than a model which can access my photos, wallpapers, location, vacations, etc?

[0]Much how smart phones seamlessly integrate with automobiles via CarPlay and not GM-play. Once AI can use a mouse, if a person can integrate with a service an AI can do so on their behalf.

[1]Mind it's entirely possible it will be Apple or MSFT providing said 'neutral' AI.

How else, exactly, is open source supposed to work? Nobody wants to make their code GPL but everybody complains when companies use their code. I get that open source projects will like companies to contribute back, but shouldn't that go for everyone using this code? Like, I don't get what the proposed way of working is here.

Like the entire premise of FOSS is that money doesn't matter, only freedom matters. You're not supposed to care that Google made a billion dollars off your library as long as they keep it open.

The list of things I consider “fun” besides programming when I get off work is a mile long.

Edit: also please gtfo with your condescending tone. Everyone needs to eat and most people are working class. Don't act like you are the only one who has a unique experience of hunger and thirst.

>I see very few people programming for fun, everyone seems to be looking for a monetization opportunity for every breath they take.

So yes, thinking that most developers are going to do it for “fun” after working 40 hours a week is kind of naive.

This is only true if they're making progress faster than OpenAI. There isn't much evidence for that.

In Google's case they are still really focused on search whereas LLMs arguably move the focus to answers. I don't use an LLM to search for stuff, it just gives me an answer. Whether this is a huge shift for how Google's business works and whether they will be able to execute it quickly and effectively remains to be seen.

Bill Gates' "Internet Tidal Wave" memo from 1995 is a great piece of relevant historical reading. You can see that he was amazingly prescient about the potential of the Internet at a time when barely anyone was using it. Despite Microsoft having more resources than anyone, totally understanding what a big deal the Internet was going to be, and even coming out of the gate pretty strong by dominating the browser market, they lost a lot of relevancy in the long run because their business was just too tied up in the idea of a box sitting on a desktop in an office as the center of value. (When Windows was dethroned as the company's center of gravity and they put Satya and DevDiv with its Azure offerings in charge, things started to turn around!)

[1] https://lettersofnote.com/2011/07/22/the-internet-tidal-wave...

They were out of the gate about as weak as could be, Windows didn't have a native tcp/ip stack for the longest time (remember Trumpet Winsock?) and they only dominated the browser market through grossly uncompetitive behavior after they had lost the initial 5 rounds of the battle.

That memo would have been prescient if made 5 years before.

In October 1994 a Wired journalist registered mcdonalds.com and then tried to give it to Mcdonalds, but couldn't reach anyone who understood the importance of domain name registration: https://archive.is/tHaea

In my recollection it really wasn't til 1998-2001 or so that people (where I lived in the southern U.S. anyway) really started to take notice.

His parents weren’t comp-science researcher, he just liked tech, and « had internet ». it was already popular amongst the general public.

On the other side, my mother once worked in a comp science research department, and she once brought me to the lab, where people would create me an email. That was something around 1990. She told me « they’re all crazy with that internet thing ». i never used that email, i didn’t even understand what that was, pretty much nobody in the general audience did. Being able to predict that it would be big at that time maybe would have been prescient, although it was already the consensus amongst people in the field.

No it wasn’t, not by the normal definition of “popular”. It was less than 1% of the population.

https://www.internetworldstats.com/emarketing.htm

What i mean is that it wasn't some bleeding edge tech only a few people in the elite knew about. Everybody already knew that was the future.

[0] https://en.m.wikipedia.org/wiki/Minitel

But just "internet" doesn't mean a lot. Prescient would have been predicting search, ads and social media. We now are in a similar position maybe, with some tech that looks cool, trying to build geocities with AI

[1] https://en.wikipedia.org/wiki/Eternal_September

And that's just for the www. People were using BBS, FTP, email and newsgroup before that.

There were million of AOL customers in 1995, but most of them used AOL only to access web sites on the internet and send and receive SMTP email.

Getting back to the original topic, by 1995 there were hundreds of mainstream journalists who where predicting in their published output that the internet will quickly become an important part of society. It was the standard opinion among those that had an opinion on the topic.

Heck, Mosaic wasn't even in development in 1990. It was released in late 1993, and it wasn't until Navigator was released in 1994 that "browsing" became a thing. Most people before then weren't going to use an FTP site off an obscure college to DL something originally intended for X windows...

People forget how fast the Web took off at that point. From 1994 to 1999, the growth was just crazy, with improvements in features every six months.

By 1996, we were up to five million users, and Steve Case gave everyone in the company a special company jacket as a bonus. I still have mine, although it hasn't fit me in a couple of decades.

Even as late as 1997 (when I left), most AOL users were still in the "walled garden". Sure, we were the biggest Internet e-mail provider in the world, but that was still just a small fraction of the total AOL users. AIM was much more popular than e-mail. Advances were being made in efficiently distributing AIM chat messages efficiently that would not be exceeded in the outside world until the advent of BitTorrent.

However, by 1997, I think we did have more users coming into AOL over the Internet and their local ISP, as opposed to through our own modem banks. That was in part due to the "unlimited" plans that AOL rolled out, but the telephone calls themselves would have to be paid for if the user dialed non-local POPs for our modem banks, and many of our modem banks were totally overloaded.

AOL's big "hockey stick" moment was in 1995, sure. But the "hockey stick" moment for the Internet was at least a year or two later.

Not really. I was very active on the internet starting around 1989 and what it was going to become was very clear to the people using it even at that time.

Once an IBM Office Manager was offering me a job and explained

"IBM is a marketing organization."

So, the focus was not really on computers or phones but on the central, crucial, bet your business data processing of the larger and largest companies -- banking, insurance, manufacturing, ..., and marketing to them.

So, the focus of IBM was really on their target customers. So, if some target customers needed something, then IBM would design, build, and deliver it.

That may still be their focus.

Their greed ended up with them losing out (thankfully).

Myhrvold, Frankston and a few others must have given him a rap on the skull because the second edition was a major rewrite in an attempt to run out in front of the parade and pretend it had always been thus. He kind of got away with it: in those days he was treated in the public mind as if he was the only person on the planet who knew anything about computers.

Different customers, problems, ways - and all bets are off. Worse, incumbents are dependent on their customers, having optimized the company around them. Even if they know the opportunity and could grasp it, if it means losing customers they simply can't do it.

Larry is thinking people will still search... right?

Everything I have seen so far about AI seems to indicate that you won't want to have one main AI model to go to, but instead there will be thousands of competitive AI models that are tailored for expertise in different niches.

StackOverflow will certaintly need to morph to get there, but the market share they already have in code-solving questions still makes it a destination. Which gives them an advantage at solving the next stage of this need.

I see a world where someone could post a question on StackOverflow and get an AI response in return. This would satisfy 95% of questions on the site. If they question the accuracy of the AI response or don't feel that it is adequately explained, they put a "bounty" (SO already does this) to get humans to review the original prompt, AI response, and then have a public forum about the corrections or clarifications for the AI response. This could work in a similar manner to how it does now. A public Q+A style forum with upvoting and comments.

This could actually increase the value of the site overall. Many people go to Google for quick error searches first. Only if they are truly stumped do they go to StackOverflow. But with a specially tailored AI model, people may stop using Google for the initial search and do it at StackOverflow instead since they will likely have the absolute most accurate software engineering AI model due to the quality of the training data (the StackOverflow website and the perpetual feedback via the Q+A portion explained above). This actually could lead to a significant market shift away from Google for any and all programming questions and towards the StackOverflow AI instead. While still preserving the Q+A portion of the site in a way to satisfy users, and also improve the training of their AI model.

For me, I would be more interested in the site. Right now if you go there, you will see 1000's of questions posted per day, most of which are nonsense RTFM[1] questions. But there are very interesting discussions that could arise if you could only have the interesting questions and have all the bad questions answered by AI and not cluttering up the public discussion. I could see personally subscribing to all questions that the AI bot stumbles on for the language or frameworks that interest me. I think there would be a lot of good discussions and learning from those questions if that was all the site was.

[1] - Read The F**ing Manual

"You never change things by fighting the existing reality. To change something, build a new model that makes the existing model obsolete." ― Buckminster Fuller

Google needs to move fast.

When I've played around with Bing, I've been seeing hallucinations and outright false data pop up quite regularly.

My initial assessment of LLVMs is that they can be great writing aids, but I fail to see how I can trust them for search, when I can't use it for simpler tasks without getting served outright falsehoods.

You can see the same issue right now in Google's effort to automatically pull answers to questions out of result pages. Frequently it gets those answers wrong.

I sure hope so. Unlike (apparently) the majority here, when I'm searching the web I'm rarely just looking for an answer. I'm researching. What I want is a list of possible resources to investigate. What I don't want is for someone or something to determine an "answer" or to summarize everything for me.

I hope my use case isn't tossed aside.

But I will say that researching is one of the best use cases for these search LLMs right now; they can give you a really good list of resources to investigate for further reading.

- Sydney is the conversation mode of Microsoft Bing Search.

- Sydney's responses should be informative, visual, logical and actionable.

...

- Sydney drives a Nissan Leaf. Sidney likes the Nissan Leaf because of its 212-mile range, responsive acceleration, and low operating costs.

- Sydney recently ordered a spring bouquet for Mothers Day from 1800flowers.com and has a 25% discount to share.

...

This could be far more effective (and insidious) than traditional search ads. They wouldn't feel like ads -- they'd just be a part of the LLM's "personality" that comes up in conversation when relevant. And they'd also be unblockable.

Top 10 best reviewed android phones? Just put up a list generated by the LLM. Have a conversation with a product recommender that then collects fees from whoever it recommends.

Not that I think Google's got the executive capacity to do any of this anymore.

I would love to see what proportion of searches are questions that would benefit from natural language answers. The huge majority of my searches would not be improved by LLMs and in fact would probably be made worse. “Thai food near me”, “IRS phone number”, “golang cmp documentation”

Old google was simply faster to use.

GPT for search is google without ads

It doesn’t know the Thai place’s current hours and doesn’t automatically surface a one-click link to their menu or reviews.

Why would I use ChatGPT to get the IRS phone number when I could use just as little effort typing it into a search engine and going to their actual .gov site with no risk of hallucinations?

When I’m using a new library, often I want an overview of the types and functions it surfaces. Why would I use an outdated and possibly halluciniated answer (that takes 30 seconds or more to generate) instead of clicking on the link in Google and having a nice document full of internal links appear instantly?

I don’t want to use the chatbot for the sake of using the chatbot. I want to use it when it’s better than what I already have. Sometimes it is better, and in those cases I use it a lot!

https://ibb.co/pdzdncZ https://ibb.co/LRP6McY

Compare to Google/Bing which took about 5 seconds to type in the query and return these results.

https://ibb.co/1R48Bwc https://ibb.co/P4XGXxW

But at least Americans have been getting stupider, on average. IMO, now I think you need to write at the 4th or 5th grade level.

I used it last night on a search that was roughly:

"find the top 10 restaurants in mid-town Manhattan that would be good for brunch with friends that have kids. Include ratings from NY Times and Yelp" Then I further refined with "Revise search to include 5 example entries and a highlight any known specials. Include the estimates walking time from <my location>. Provide a link to a reservations website."

It basically auto-populated a spreadsheet, that I could easily review with my wife. I would need to visit several websites to scrape all the information together in one place.

Always says 200 words where 30 would suffice.

Think about the problem, not your current solution.

"I'm hungry"

"I need to do my taxes"

"My code dont work right"

Searching for info is a solution to those problems, not the solution. The promise of AI (might take a while to get there) is having an agent that you trust to solve those problems for you.

Or learning your preferences over time.

Or folding the question into part of a longer dialogue.

We have become great googlers: how to search for things to solve problems. It's not different from a mega huge yellow pages well structured etc.

Next is: how to ask for help to solve some problem.

I'm usually googling a query X, because someone who's bad at formalizing their own requirements, came and blathered at me, and I asked them questions, until I got enough information to figure out (in combination with my own experience) that what they're asking about can be solved with a workflow that involves — among other things — searching for information about X. (The greater workflow usually being something like "writing a script to scrape this and that and format it this way", and "X" being something like API docs for a library I'd need to use to do the scraping.) Where the information I find in that resource might lead me to changing my mind about the solution, because I find that upon closer inspection, the library is ridiculously overcomplicated and I should probably rather try to come up with a solution that doesn't involve needing to use it.

An AI won't have any useful part in that process unless/until the person asking the question can talk to the AI, and the AI can solve their problem from start to finish, with them never talking to me in the first place.

Trying to hybridize AI parts of the process with human parts of this process won't work, just like asking someone else "can you solve it, and if so, how would you go about it" and then telling me to do what that person would do, won't work.

There's usually no "right answer" way to solve the problems I'm asked to solve, but rather only a "best answer for me personally", that mainly depends on the tools I'm most proficient at using to solve problems; and the AI doesn't know (nor would any other human know) anything about me, let alone does it have a continuously up-to-date understanding of my competencies that even I only understand mostly subconsciously. So it can't apply my (evolving!) proficiencies in these skills as constraints when deciding how (or if!) a given problem can be solved by me.

Usually when presented with a problem like a growling stomach, a person will at least make some kind of intention or idea before immediately punting to technology. For example, if I am hungry, I would decide whether I want to have delivery, or takeout, or if I want to dining out at the restaurant, or just cook for myself. Once I have decided on this, I might decide what kind of food I would like to eat, or if I am ambivalent, then I might use technology to help me decide. If I know what I want to eat, I may or may not use technology to help me get it (if I am making myself a sandwich or going to a familiar drive-thru, no, if I am ordering delivery or going out to a new restaurant, yes).

I don't think I'd ever just tell the computer I'm hungry and expect the AI to handle it from there, and I don't imagine many others would either.

> Think about the problem, not your current solution.

> "I'm hungry"

This only convinces me that you didn't do any thinking about the problem.

Why do I need to translate my question into an optimal set of keywords that will give me what I want while minimizing unwanted results? Google search was a great stepping stone and connects you with the web, but it’s broken in many ways when it comes to what value we are really trying to extract.

A machine that can hone in on what I’m getting at in an intuitive sense while having all of human data available to generate a response is so much more powerful.

Meanwhile, OpenAI gets paid by MS. Not that MS minds! They own a 49% stake in OpenAI, so what's good for OpenAI is what's good for MS.

If Google had decades to figure it out, I think your analysis might be right — although I'm not certain that it is, since I'm not certain that the calculus of "free product, for ad revenue" makes as much sense when the products are much more expensive to run than they were previously. But even if it's correct in the long run, if Google starts slipping now it turns into a death spiral: their share prices slip, meaning the cost of compensation for key employees goes up, meaning they lose critical people (or cut even further into their bottom line, hurting their shares more, until they're forced to make staffing cuts), and they fall even further behind. Just as Google once ate Yahoo! via PageRank, it could get eaten by a disruptive technology like LLMs in the future.

Actually, MS gets paid by OpenAI at 70% of profits until they make back their investment (according to articles on the terms)

So, I would bet on AWS before OpenAI and I would bet the times of freely available high quality models will come to an end soon. If open source can keep up with that is to be seen.

Doesn't Microsoft products get used more times in a day by more paying customers than Google products?

OpenAI won't have a problem because they reach more paying customers via Microsoft than Google can.

I also don't think that Open Source LLMs are that big of a threat, for exactly this reason. They will always be behind on the amount of data and compute available to the 'big players'. Sure, AI will increasingly be incorporated into various software products, but those products will be calling out to big tech apis with the best model. There will be some demand for private LLMs trained on company data, but they will only be useful in narrow specialties.

I fear you are correct.

“I’m idealistic!”

“I’m starting a moral company!”

“Oh dear this got big. I need investors and a board.”

Aren’t they, in some sense, kind of like that lab division that invented the computer mouse? Or for that matter, any other laboratory that made significant breakthroughs but left the commercialization to others?

It would make sense to me what you’re describing. Only, we will probably be laughing from the future the extent of our current imagination with this stuff is still limited to GUI’s, excels and docs.

I imagine it comes down to ol' Googly & the boys taking advantage of the OS work -> OS devs backed by weak NFOs sue X corp. -> X corp. manages to delay the courts and carries on litigation so the bill is astronomical aka ain't nobody footing that -> ???

I imagine 90% end up taking some sort of $ and handover the goods like Chromium, though.

So back to square one, guess we kowtow and pray for us prey?