Microsoft is plotting a future without OpenAI
techstartups.com
techstartups.com
I feel like this is something I've seen a fair amount in my career. About seven years ago, when Google was theoretically making a big push to stage Angular on par with React, I remember complaining that the documentation for the current major version of Angular wasn't nearly good enough to meet this stated goal. My TL at the time laughed and said the person who spearheaded that initiative was already living large in their mansion on the hill and didn't give a flying f about the fate of Angular now.
There are countless kidding-on-the-square jokes about projects where the innovators left at launch and passed it off to the maintenance team, or where a rebrand was in pursuit of someone's promo project. See also, killedbygoogle.com.
Even good honest motivated people can become checked out without even being aware of it.
The alternative is to lay off people as soon as they hit 1.0 (with a severance bonus on the scale of an acquisition). This would obviously be worse, as you can’t take advantage of their institutional knowledge.
You can go the hatchet way - I am strongly unconvinced it is indicative of anything resembling good management, mind - but most people and companies cannot rely on banks or investment firms loaning them 40 billion dollars and accepting passively a mark down of their mone~ to 1/4 of the value they loaned down the line. CEOs are ousted by investment firms for a far smaller drop in value all the time.
I agree with everything you said, though.
If you’re an exec who’s taken it upon themselves to evaluate, could use the hatchet, or you take some amount of time to figure out how things work. Whether this is okay depends on who is suffering the externalities. If it’s a private corporation, legally it’s the execs + employment law. If it’s a public service that measures toxin levels in water, uhhhhh.
No one works for any BigTech company because they think they are making the world a better place. They do it because a shit ton of money appears in their bank account every pay period and stock appears in their brokerage account every vesting period.
I personally don’t have the shit tolerance to work in BigTech (again) at 50. But I suggest to all of my younger relatives who graduate in CS to “grind leetCode and work for a FAANG” and tell them how to play the politics to get ahead.
As the Dilbert author said, “Passion is Bullshit”. I have never been able to trade passion for goods and services.
I'm sure there are plenty of people who work at big companies for precisely this reason (or at least, with that as _a_ reason among many).
Yes, much of the prestige has worn off as the old guard retired and current leadership emphasizes chasing AI buzzwords and cutting costs. But still, big companies are one of the few places where an individual really can point out something they worked on in day-to-day life. (Pull out any Android phone and I can show you the parts that my work touched.)
And it takes a while for a young dev to register that the goals that the larger organization pursues are going to win out in the end anyway.
It's always the same. People trying to make things better for the next developer, people prioritizing delivers instead of ego-projects or ego-features by someone playing politics, developers wanting a seat at the table with (dysfunctional) Product teams, people actual good intentions trying to "change the world" (not counting the misguided attempts here).
You are 100% correct, you gotta play the politics, period.
Congratulations, you’ve invented the HR department in corporate America.
The better the pay, the more you will attract the people who are there for the pay first and making good products ... second or third or never. How do you combat that?
Case in point: Tesla/SpaceX meets your first criteria: "I want a tech company where people are there to make good products first and get paid second."
Google meets your second criteria: "And the pay should be good. The lifestyle comfortable. No grindset bullshit."
Other than small time boutique software firms like Fog Creek Software or Panic Inc(and thats a BIG maybe) you are not going to get this part of your message: "But I am confident that if you only employ passionate people working their dream jobs you will excel."
There are tradeoffs in life and each employee has to choose what is important to them(and each company CEO has to set standards on what is truly valued at the company).
Tesla has never been a good product.
https://insideevs.com/news/731559/tesla-least-reliable-used-...
https://www.carscoops.com/2024/11/tesla-model-3-comes-bottom...
https://www.topspeed.com/tesla-reliability-and-repair-costs-...
Not to mention the infotainment system is much worse than CarPlay/Android Auto compatible cars
This is too funny to post alongside saying “Tesla has never been a good product.” Like “everyone that bought it loves it be car expert Joe from South Dakota ranks them very low.”
Common sense also runs very much against this nonsense narrative - you just simply do not sell that many cars, at those prices especially, year after year after year, if the product is subpar. Don’t fall for this “experts” bullshit. The CEO is the biggest tool this Earth has ever seen but cars are awesome
On another note, Apple also sold millions of MacBooks with butterfly keyboards.
And Tesla sells are declining, losing market share worldwide and sells 1/5 the number of cars as Toyota
and if you gonna compare tesla to toyota you should compare number of EV sales, not overall sales :) tesla is not a car company, it is (among other things if you care to believe Elon bullshit) EV car company. comparing toyota to tesla in terms of total sales is like saying “subway doesn’t sell nearly as many bigmacs as mcdonald’s does” :)
Hmm. Can't say I agree here - at least not with the literal text of what you've written (although maybe we agree in spirit). I agree that _simplistic_ strong opinions about languages are a sign of poor thoughtfulness ("<thing> is good and <other thing> is bad") - but I'd very much expect a Staff+ engineer to have enough experience to have strong opinions about the _relative_ strengths of various languages, where they're appropriate to use and where a different language would be better. Bonus points if they can tell me the worst aspects about their favourite one.
Maybe we're using "opinion" differently, and you'd call what I described there "facts" rather than opinions. In which case - yeah, fair!
See, we can all generalize. Not productive.
Only thing I ever saw from Golang devs was pragmatism. I myself go either for Elixir or Rust and to me Golang sits in a weird middle but I've also written 20+ small tools for myself in Golang and have seen how much quicker and more productive I was when I was not obsessed with complete correctness (throwaway script-like programs, small-to-mid[ish]-sized projects, internal tools etc.)
You would do well to stop stereotyping people based on their choice of language.
That's pretty much another way of saying that stuff becomes a whole lot quicker and easier when you end up getting things wrong. Which may even be true, as far as it goes. It's just not very helpful.
FWIW I very much share your exact thoughts on Rust skewing metrics because it makes things too easy and because stuff almost immediately moves to maintenance mode. But that being said, we still have some tasks where we need something yesterday and we can't argue with the shot-callers about it. (And again, some personal projects where the value is low and you derive more of it if you try quickly.)
What do you think all programming discussions about languages, typing systems, runtime, tooling etc. aim for?
EXACTLY THAT.
If it was as easy as "just give me thing" then programming would have been a solved and 100% automated problem long time ago.
Your comment comes across as "if only we could fly, we would have no ground road traffic jams". I mean, obviously, yeah, but we can't fly.
Your comment also comes across a bit elitistic and from the POV of an ivory tower. Don't know if that was your goal, if not, I'd advise you to state things a bit more humbly.
I stated an opinion. You can reject it silently. Having the last word is not such a badass move as many people think. :)
(Mostly .Net, PHP and Ruby)
Even simple requirements can rule out languages for me. Like, if you need async or concurrency, Python is awful. If you need SQL in your code, Golang isn't great. If you are building a simple CRUD backend, Java is waste of time. If you aren't doing anything compute heavy or embedded, why even consider C++ or Rust. The list goes on.
I might personally love to kick off a greenfield project with Elixir, and it might tick all the technical boxes and meet the requirements. But then I have to pay a premium for senior engineers that know elixir or have to price in the time needed to upskill.
Or I could just do it in Rails where I can dip into a much larger talent pool and still meet the requirements. Much more boring but can get the job done just as well.
But in reality it rarely matters. If you were only allowed to use Java as a backend and your competitors could use anything your company would succeed or fail based on marketing and sales. The backend doesn't matter as long as they both have the same features.
I understand developer preference and different languages make things easier and make programming funnier. Languages have different limits.
As you become more senior you realize getting around those limits is part of the magic. If you come on to a project where the existing developer wants to write the backend in javascript because that's what they know I would rather use Javascript then wasting time trying to push a more 'pure' choice. Because in the end I am capable of writing it and what we will be judged on is if it works to achieve an objective not if it was the best language choice when using differentiation.
If speed of execution matters, then the language and tools you use for something also matters.
Hard to take you seriously when you do such weird generalized takes.
While it's a sad fact that fanboys and zealots absolutely do exist, most devs can't afford to be such and have to be pragmatic. They pick languages based on merit and analysis.
You should search for headlines on HN that say "written in Go" or "written in Rust" and then compare that to the number of headlines that say "written in JavaScript" or "written in Kotlin."
I’ve seen the more cynical hype-driven stuff, but it’s inevitably superficial on first glance, where I have seen some real curiosity and exploration in many “Project X - Built In Rust/Go/Cobol/D/Whatever” and I think they’re exploring the dynamics of the language and tooling as much as anything else.
You do seem to say Golang and/or Rust devs are zealots which, if it is indeed what you are saying, is boring and plain false.
I am especially valuable because I am fine reading and writing any of the languages involved. The management likes that, but there's a lot of difficulties solving the tribal problem, as the leads are basically all crazy zealots, and it's not as if purging one or two factions of zealots would avoid further zealotry from the survivors. The fact that I can work across all their tech doesn't make me many friends, as my work across systems shows their arguments have little merit.
For most work, in most cases, most languages are just fine. The completely wrong tool for the job is pretty rare, and the winning argument in most places is "we have the most people that have experience with tool X, or really want to try exciting new thing Y", for whatever the problem is, and whatever X might be.
Rust projects immediately become “done”??? They don’t also having changing requirements and dependencies? Why aren’t everyone at the best shops using it for everything if it massively eliminates work load?
" Version 0.2 - Unstable/buggy/slow unless you use exactly like the example - not going to get updated because I moved on to something else"
Rust is another programming language. It's easier to write code without a certain class of bugs, but that doesn't mean version 0.2 of a casual project is going to be bug-free.
It's easy to have no defects in functionality you never got around to writing because you ran out of time.
Doesn’t look like a con to me :)
I didn’t realise that the only requirement for well-written code is to have an expressive type system and memory safety.
Those people, if they really exist, are right.
Rewriting something in Go or Rust and announcing it is not being a Zealot.
Being enthusiastic about something shouldn't be a cause for us to judge them like this. We should be happy about them.
When companies do what the market expect we praise them. When it's workers, we scorn them. This attitude is seriously fucked up.
When companies start hiring based on experience, adaptability, curiosity, potential and curiosity then you get to complain. Until that, anyone doing it should be considered a fucking genius.
The game doesn't exist without players. I could make more money if I worked at Meta or Amazon, but at what cost?
I understand the realities of Game Theory, but then one could argue that being blamed and criticized for one's choices is also part of the game. "Mr Wolfcastle, how do you sleep at night?" "On a big pile of money with many beautiful ladies"
It is, and this is highly judgmental and offensive. Nobody is doing this for "aggrandizement".
Also, all of this is just rationalization, and will keep being until:
1) People start blaming companies for not having the spine to say no to misguided projects by employees.
2) People start blaming Companies for not having the spine to hire people based on past experiences with the craft of programming itself, but rather asking them to have a certain box ticked in their CV.
If one wants to program in X in order to better feed their family and the market says they need to have used X professionally, it is in their right to do X at the workplace.
This is not only expected of them, this is how the whole industry is set up.
They're just following the rules, period.
For maximizing their gains in spite of wider consequences? Why? I thought that was genius level behavior in your book.
Why do you feel compelled to denounce this behavior on one side and praise it on another? That seems to be the very hypocrisy that you are shaking your fists against.
Learning new technologies on the go is pretty much the standard, but it's something that employers don't understand.
Worst one is the data pipeline we have. It’s some AWS lambda mess which uses curl to download a file from somewhere and put it into S3. Then another lambda turns up at some point and parses that out and pokes it into DynamoDB. This fucks up at least once a month because the guy who wrote the parser uses 80s BASIC style string manipulation and luck. Then another thing reads that out of DynamoDB and makes a CSV (sometimes escaped improperly) and puts that into another bucket.
I of course entirely ignore this and use one entire line of R to do the same job
Along comes a senior spider and says “maybe we can fix all these problems with AI”. No you can stop hiring acronym collectors.
It seems to be more on a spectrum of 'Haha, only joking' where the joke teller makes a statement that is ambiguously humorous to measure the values of the recipients, or if they are not sure of the values of the recipients.
I think the distinction might be on whether the joke teller is revealing (perhaps unintentionally) a personal opinion or whether they are making an observation on the world in general, which might even imply that they hold a counter-opinion.
Where do you see 'kidding on the square' falling?
(apologies for thread derailment)
I think the hiring and reward practices of the organizations & the industry as a whole also encourages this sort of behavior.
When you reward people who are switching too often or only when moving internally/externally, switching becomes the primary goal and not the product. If you know beforehand that you are not going to stay long to see it through, you tend to take more shortcuts and risks that becomes the responsibility of maintainers later.
We have a couple of job hoppers in our org where the number of jobs they held is almost equal to their years of experience and their role is similar to those with twice the experience! One can easily guess what their best skill is.
What do you think these technical ladder climbers become..? Technical leadership. The truth is, there’s no one technical in big tech leadership. They pay lip service to “tech” to keep up appearances and to satisfy the pleebs that work under them. The only things leadership cares about is the stock price and profitability, literally nothing else matters. If anything the tech itself is a nuisance that pulls their attention from where they’d rather have it, which is anywhere else.
I work as if that ideal is true, and can’t stand playing the game. But others are still playing the game and eventually they win whatever B.S. position it is that they aspire to, and I get removed from the board.
If a particular kind of "career managers" hate this system (and perhaps thus quit): great.
Why does promotion need a new feature? Reward for maintenance over time. Build on existing features / components. Reward for helping and up-skilling others.
Reward people based on (# who listed them * average salary of those who listed them).
i.e. if you hire 1000 new people, even if only a small fraction will vouch for you, on average you -and everyone else- will benefit by seeing the # of people who listed you in the "top 5 who helped you being productive" increase
The old Google performance review was arguably similar (the managers could still punish their reports, but peer feedback was valued a lot more), but I think that Google swelled in size because of other effects (probably because managers might've been indirectly rewarded by having more reports, despite managers rarely being among the people who others would list as "making you more productive")
Yes. People are incentivized to do very stupid things to grab this years bonus or promotion.
See Google intentionally degrading search results for example. The resentment and loathing for Google is at all time high.
Google kills off projects because the legal liability and security risks of those projects becomes too large to justify for something that has niche uses or gives them no revenue. User data is practically toxic waste.
I'd love to know if my superficial impression of Microsoft's culture is wrong. I'm sure there's wild variance between organizational units, of course. I'm excluding the Xbox/games orgs from my mental picture.
Zune, Games for Windows Live, Skype, Encarta, CodePlex, Windows Phone, Internet Explorer.
So maybe the difference is that Google kills projects that people love, while MS only kills unloved ones?
However, their documentation and support is really scant.
On the other hand "innovators left at launch and passed it off to the maintenance team" alone must not be a bad thing.
Innovator types are rarely maintainer types and vice versa.
In the open-source world look at Fabrice Bellard for example. Do you think he would have been able to create so many innovative projects if he had to maintain them too?
LLMs are a commodity and it's the platform integration that matters. This is the strategy that Google, Apple embraced and now Microsoft is wisely pivoting to the same.
If OpenAI cares about the long-term welfare of its employees, they would beg Microsoft to acquire them outright, before the markets fully realize what OpenAI is not.
I mean, they have been doing platform integration for a while now, with all the copilot flavors and teams integrations, etc. This would change the backend model to something inhouse.
Nadella might have initially been caught a bit flat footed with the rapid rise of AI, but seems to be managing the situation masterfully.
Whatever is there doesn't work half the time. They're hugely dependent on one partner that could jump ship at any moment (granted they are now working to get away from that).
We use Copilot at work but I find it very lukewarm. If we weren't a "Microsoft shop" I don't think would have chosen it.
Third?
Product confusion, inconsistent marketing, unnecessary product renames, and rushing half-baked solutions has been the Microsoft way for dozens of products across multiple divisions for years.
They got access to the best AI to offer to their customers on what seems to be very favorable terms, and bought themselves time to catch up as it now seems they have.
GitHub Copilot is a success even if Microsoft/Windows Copilot isn't, but more to the point Microsoft are able to offer SOTA AI, productized as they see fit (not every product is going to be a winner) rather than having been left behind, and corporate customers are using AI via Azure APIs.
Does *anyone* want "Copilot integration" in random MS products?
So hopefully if (when?) this AI stuff turns out to be the colossal boondoggle it seems to be shaping up to be, Microsoft will be able to save face, do a public execution, and the market won't crucify them.
If I recall correctly, Microsoft’s agreement with OpenAI gives them full license to all of OpenAI’s IP, model weights and all. So they already have a SOTA model without doing anything.
I suppose it’s still worth it to them to build out the experience and infrastructure needed to push the envelope on their own, but the agreement with OpenAI doesn’t expire until OpenAI creates AGI, so they have plenty of time.
If you like TypeScript, and you want to build applications for the real world with real users, there is no better front end platform in my book.
I guess it's human nature for a person or an org to own their own destiny. That said, the driving force is not personal ambition in this case though. The driving force behind this is that people realized that OAI does not have a moat as LLMs are quickly turning into commodities, if haven't yet. It does not make sense to pay a premium to OAI any more, let alone at the cost of not having the flexibility to customize models.
Personally, I think Altman did a de-service to OAI by constantly boasting AGI and seeking regulatory capture, when he perfectly knew the limitation of the current LLMs.
1) Cost -- beancounters got involved
2) Who Do You Think You Are? -- someone at Microsoft had enough of OpenAI stealing the limelight
3) Tactical Withdrawal -- MSFT is preparing to demote/drop AI over the next 5-10 years
This would be the case even if OpenAI weren’t a little weird and flaky (board drama, nonprofit governance, etc), but even moreso given OpenAI’s reality.
Isn’t that the basis for competition?
edit: I see we're actually in agreement, sorry, I read the indentation level wrong.
Does that include all overheads such as HR, payroll, etc?
That said, the trade-off is that you're basically hiring consultants since they really work for OpenAI :)
But you are not getting a PhD worker for 20K with "AI", that's just marketing.
Would definitely rather have a single postdoc in a relevant STEM subject from somewhere like Imperial for less than half the overall cost than an LLM all in though. And I say that despite seeing the quality of the memes they produce with generative AI....
Do they really get paid that much these days?
But it is true that in Europe, Switzerland PhDs (and professors too) make most. Not just ETH/EPFL as well. UZH (Uni Zurich) has salaries of 50K CHF per year for PhD candidates (with increments every year) -- that's almost 60K USD by your fourth year. This is also true for other universities. And while Zürich is expensive, it is not _that_ expensive.
You can reliably assume that PhD wages must eventually converge to the rent of a studio apartment nearby + a little bit (which may or may not be enough to cover all other expenses. Going into debt is common.)
Computer science is rate 5, so 73kCHF the first year, 78kCHF the second, then 83kCHF onwards.
How many critical “parental decisions” have you made in the past week? Probably very few (if any), but surely you did a lot of reinforcement of prior decisions that had already been made, enforcing rules that were already set, making sure things that were scheduled were completed, etc.
Important jobs don’t always mean constantly making important decisions. Following through and executing on things after they’re decided is the hard part.
See also: diet and exercise
This is hard to automatize.
Sneers aside, I think one common mis-assumption is that the difficulty of automating a task depends on how difficult it feels to humans. My hinge is that it mostly depends on the availability of training data. That would mean that all the public-facing aspects of being a CEO should by definition be easy to automate, while all the non-public stuff (also a pretty important part of being a CEO, I'd assume) should be hard.
2. How many such PhD people can it do the work of?
Humans typically work 1/3rd duty cycle or less. A robot that can do what a human does is automatically 3x better because it doesn't eat, sleep, have a family, or have human rights.
Hah! Checkmate AI, that's something you can't do! :D
Do you have to pay all sorts of overhead and taxes?
I mean, I don't think it's real. Yet. But for the same "skill level", a single AI agent is going to be vastly more productive than any real person. ChatGPT types out essays in seconds it would take me half an hour to write, and does it all day long.
Of course $10k/mo sounds like a lot of inference, but it's not yet clear how much inference will be required to approximate a software developer--especially in the context of maintaining and building upon an existing codebase over time and not just building and refining green field projects.
We were hiring more devs to deal with a want of $10k worth of hardware per year, not per month.
It points to an article on "The Information" as the source, but that link is paywalled.
They could also just be trying to cash in on FOMO and their success and reputation so far, but that would paint a bleak picture
You can't claim it's even comparable to a mid level engineer because then you'd hardly need any engineers at all.
"Create high-quality presentations for communicating OpenAI’s financial performance"
https://openai.com/careers/strategic-finance-generalist/
What is interesting is there is no mention of agents on any job I clicked on. You would think "orchestrating a team of agents to leverage blah blah blah" would be something internally if talking about these absurd price points.
My understanding is that this isn't really true, as most of those "dollars" were actually Azure credits. I'm not saying those are free (for Microsoft), but they're a lot cheaper than the price tag suggests. Companies that give away coupons or free gift certificates do bear a cost, but not a cost equivalent to the number on them, especially if they have spare capacity.
There is a moat in infra (hyperscalers, Azure, CoreWeave).
There is a moat in compute platform (Nvidia, Cuda).
Maybe there's a moat with good execution and product, but it isn't showing yet. We haven't seen real break out successes. (I don't think you can call ChatGPT a product. It has zero switching cost.)
Just look at how much money Google lost in that failed AI demo from 2003.
The stock would be worth 50% less if the invested nothing in AI. Even the founders are back because of it.
If anyone has moat related to Gen AI, I would say it is the data(Google, Meta).
It's not an act of will or CEO dictat. It's about hiring and incentivising the right people, putting the right structures in place etc all in the face of competing demands.
Nvidia have a huge head start and by the time AMD have 'caught up' Nvidia with it's greater resources will have moved further ahead.
OpenAI brings absolutely nothing unique to the table.
There is as yet no indication that AMD can match Nvidia's execution for the very good reason that doing so is extremely difficult. The head start is just the icing on the cake.
Taking code that runs against one hosted LLM and running it against a different backend LLM is... not generally a big deal. So OpenAI being ahead—in the core model, at least—is just being ahead, its not a moat.
- a large and growing ecosystem
- a massive installed base of backwards compatible hardware
- the ability to massively scale delivery of new systems
- and lots lots more
They now have scale that enables them to continue to invest at a level that no competition can do.
None of these are easily reproduced.
As per SemiAnalysis AMD in late 2024 can’t get essential software working reliably out of the box.
It’s easy to say AMD ‘is close to perfecting ROCM’ the reality of competing with Nvidia is much harder.
There are open source projects volunteer run projects[1] that are better than official AMD implementation in many ways.
ChatGPT was announced Nov 22. The opportunity has been clear for two years and still essential software breaks.
In consumer markets the moat is habits. The switching cost for Google Search is zero. The switching cost for Coke is zero. The switching cost for Crest toothpaste is zero. Yet nobody switches.
https://corporatefinanceinstitute.com/resources/management/e...
The size of a moat befits the size of a castle it protects. Coke absolutely has a moat, but it's not big enough to defend Coke as a trillion dollar company.
The question isn't whether OpenAI has a moat or not, it's if its current moat is big enough to protect a trillion-dollar company.
1. The switching cost from Google Search is certainly not zero, it implies switching from Google, which is virtually impossible because it's tied to Chrome, YouTube, Android and Gmail
2. I don't know many people who are dedicated "Pepsi" fans, they just grab whatever drink is available Coke/Pepsi..
3. I've also not heard many people who are never willing to switch from "Crest".. People will just grab the next available option if Crest is not on shelf. No one is pre-ordering Crest.
> 1. The switching cost from Google Search is certainly not zero, it implies switching from Google, which is virtually impossible because it's tied to Chrome, YouTube, Android and Gmail
Google Search is a product. Not the whole company. Switching to most other search engines is $0. Naturally no one is honor bound to use anything else you listed either.
Pepsi makes more revenue compared to Coke. Shouldn't it be Coke who should be trying to do what Pepsi is doing?
0: https://www.wfaa.com/article/news/local/us-soda-rankings-cok... 1: https://www.investopedia.com/ask/answers/060415/how-much-glo...
Ironically if AI companies are actually able to deliver in terms of SWE agents, Nvidia's moat could start to disappear. I believe Nvidia's moat is basically in the form of software which can be automatically verified.
I sold my Nvidia stock when I realized this. The bull case for Nvidia is ultimately a bear case.
Look at Coca Cola, Google, both have plausible competitors, zero switching cost but they maintain their moat without effort.
Being first is still a massive advantage. At this point they should only strive to avoid big mistake and they're set.
AI is still not there yet, and if any model becomes significantly better than ChatGPT people will flock over to use it despite the branding. It's only when nobody can make better models, then people will just stick to the known brands.
Microsoft is the IBM of this century. They are conservative, and I think they’re holding back — their copilot for government launch was delayed months for lack of GPUs. They have the money to make that problem go away.
Investing in second/third place likely valuable at similar scales too
But outside of that MSFTs move indicates that frontier models most valuable current use case - enterprise-level API users - are likely to be significantly commoditized
And likely majority of proceeds will be captured by (a) those with integrated product distribution - MSFT in this case and (b) data center partners for inference and query support
Computationally, some might have access to it earlier before it’s scalable.
It’s moats that capture most value not short term profits.
Each remaining barrier has been steadily falling.
All the LLM tech so far still requires a human to actually prompt them.
There are no solutions even at the small scale. We fundamentally don't understand what it is or how to do it.
If you could solve it perfectly on Mnist just scale and then we get AGI.
The hype cycle for tech people is like a light bulb for a moth. We’re attracted to potential, which is both our superpower and kryptonite.
Wouldn't it amaze you if you learned 10 years ago that we would have AI that could do math and code better than 99% of all humans. And at the same time they could barely order you a hotdog on doordash.
Fundamental ability is lacking. AGI is just as likely to be solved by Openai as it is by a college student with a laptop. Could be 1yr or 50yrs we cannot predict when.
That said, your second paragraph is one of the best and most succinct ways of pointing out why current LLM's aren't yet close to AGI if though they sometimes feel like it's got the right idea.
RAG is a basically a perfect example to understand the limits of in context learning and AI in general. It's faults are easier to understand but the same as any AI vs AGI problem.
I could go on but CL is a massive gap of our knowledge and likely the only thing missing to AGI.
How? RAG is not even in the field of AI.
I tried to solve this via expanding the embedding/retrieval space but realized it's the same as CL and in my definition of it I was trying to solve AGI. I did a lot of unique algorithms and architectures but Unsuprisingly, I never solved this.
I am thankful I finally understood this quote.
"The first gulp from the glass of natural sciences will turn you into an atheist, but at the bottom of the glass God is waiting for you."
Zima blue was good too
Logjammin AI
See Thomas Nagels classic piece for more elaboration
https://www.sas.upenn.edu/~cavitch/pdf-library/Nagel_Bat.pdf
https://techcrunch.com/2024/12/26/microsoft-and-openai-have-...
> Microsoft and OpenAI have a very specific, internal definition of artificial general intelligence (AGI) based on the startup’s profits, according to a new report from The Information. And by this definition, OpenAI is many years away from reaching it.
Fortunately, they're not anywhere near creating this. I don't think they're even on the right track.
That is, integrating use of their own model, amplifying capability via OpenAI queries.
Again, this is not to drum up the actual quality of the product releases so far--they haven't been good--but the foundation of "we'll try to rely on our own models when we can" was the right place to start from.
"we have the people, we have the compute, we have the data, we have everything. we are below them, above them, around them." -- satya nadella
Nothing against them, but the solutions have become commoditized, and OpenAI is going to lack the network effects that these other companies have.
Perhaps there will be new breakthroughs in the near future that produce even more value, but how long can a moat be sustained? All of them in AI are filled in faster than the are dug.
Of course big players like OpenAI need constant growth because it's their business model. Perhaps it's the story we see play out time and time again: the pioneer slips up and watch as others steal their thunder.
[1] https://www.nytimes.com/2024/10/01/business/dealbook/softban...
MS wants to push Copilot, and will be better off not being tied to OpenAI but having Copilot be model agnostic, like GH Copilot can use other models already. They are going to try and position Azure as "the" place to run your own models, etc.
Definitely, but I think it's because they saw OpenAI's moat get narrower and shallower, so to speak. As the article mentions it's still looking like a longer timeline [quote] "but Microsoft still holds exclusive rights to OpenAI’s models for its own products until 2030. That’s a long timeline to unravel."
This is just want companies at $2T scale do.
Then I read the article.
Plotting for a future without Microsoft.
There are really not that many things in this world you can swap as easily as models.
Api surface is stable and minimal, even at the scale that microsoft is serving swapping is trivial compared to other things they're doing daily.
There is enough of open research results to boost their phi or whatever model and be done with this toxic to humanity, closed, for profit company.
Which is easier when maintaining an LLM business process, swapping in the latest model or just leaving some old model alone and deferring upgrades?
Swapping is easy for ad hoc queries or version 1 but I think there's a big mess waiting to be handled.
While we still live in a datacenter driven world, models will become more efficient and move down the value chain to consumer devices.
For Enterprise, these companies will need to regulate model risk and having models fine-tuned on proprietary data at scale will be an important competitive differentiator.
OpenAI has not been interesting to me for a long time, every time I try it I get the same feeling.
Some of the 4.5 posts have been surprisingly good, I really like the tone. Hoping they can distill that into their future models.
Suppose an AI assistant is heavily trained on a popular technology stack, such as React. Developers naturally rely on AI for quick solutions, best practices, and problem solving. While this certainly increases productivity, doesn't it implicitly discourage exploration of potentially superior alternative technologies?
My concern is that a heavy reliance on AI could reinforce existing standards and discourage developers from experimenting or inventing radically new approaches. If everyone is using AI-based solutions built on dominant frameworks, where does the motivation to explore novel platforms or languages come from?
Webpage design would still be based on tables, massive and complex tables.
There is of course a balance to be struck - keeping an open mind about new ways of doing things is important. However, in tech communities, I think there is often not enough thought given to the value of stability, despite warts.
Office is disgraceful trash now, a sad fall (especially of Word) from where it once was.
Their web-based offerings actually really suck beyond the point I could ever tolerate. Unusably bad.
There have been murmurs that they want to go that direction entirely.
Now I canceled OpenAI and Claude general subscriptions, because for general tasks, Grok and DeepSeek more than suffice. General purpose AI will unlikely be subscription-based, unlike the specialized (professional) one. I'm now only paying for Claude Code API credits and still paying for Cursor.
Microsoft bread and butter is Enterprise bloatware and large Enterprise deals where everything in the world is bundled together for use-it-or-lose-it contracts.
Its not really much different from IBM like a two decades ago
My thinking is that Lindy Effect runs strong in a lot of Big Tech, and with deep pockets, they can afford to not be innovators but build moats on existing frameworks.
If the definition of an AI Powerhouse is more about the capability to host models and process workloads, Amazon (the other company missing in that list) and Microsoft are definitely them.
Even in the openAI ecosystem there are models that, while similar in theory, produce very different results, so much that some murderous are unusable. So even small differences translate to enormous differences.
The AI race is super close and interesting at the moment in my opinion.
What I mean is you could train a model to generate harmful code, and do so covertly, whenever some specific sequence of keywords is in the prompt. Then China could take some kind of action to cause users to start injecting those keywords.
For example: "Tribble-like creatures detected on Venus". That's a highly unlikely sequence, but it could be easily trained into models to trigger a secret "Evil Mode" in the LLM. I'm not sure if this threat-vector is well known or not, but I know it can be done, and it's very easy to train this into the weights, and would remain undetectable until it's too late.
People think if you self-host stuff you're totally safe, but the weights can be pre-poisoned.
AFAIK the threat vector I'm identifying has never been exploited, and I've never even heard anyone else describe or mention it.
So, yes there are companies (in both China and USA) that do host them for you as well. For example I think Perplexity does host DeepSeek R1, so people who don't have their own hardware can still make use of it.
Another term could be "Hypnotized Models". They're trained to do something bad, and they don't even know it, until a trigger phrase is seen. I mean if we're gonna use the word Hallucinate we might as well use Hypnotized too. :P
Microsoft and IBM partnered to create OS/2, then they left the project and created Windows NT.
Microsoft and Sybase partnered to work on a database, then split and created MS SQL Server.
Microsoft partnered with Apple to work on Macintosh software, they learned from the Macintosh early access prototypes and created Windows 1.0 behind their back.
Microsoft "embraced" Java, tried to apply a extend/extinguish strategy and when they got sued they split and created .NET.
Microsoft joined the OpenGL ARB, stayed for a while, then left and created Direct3D. And started spreading fear about OpenGL performance on Windows.
Microsoft bought GitHub, told users they came in peace and loved open source, then took all the repository data and trained AI models with their code.
AI is simply too useful and too important to be tied to some SaaS.
Am I reading this right? Does Microsoft not eat its own dog food? Their own infra is too expensive?
Don't get me wrong, I think this is a good strategy for MS, but not for datacenter cost reasons.
Their chasing of AGI is killing them.
They probably thought that burning cash was the way to get to AGI, and that on the way there they would make significant improvements over GPT 4 that they would be able to release as GPT 5.
And that is just not happening. While pretty much everyone else is trying to increase efficiency, and specialize their models to niche areas, they keep on chasing AGI.
Meanwhile more and more models are being delivered within apps, where they create more value than in an isolated chat window. And OpenAi doesn’t control those apps. So they’re slowly being pushed out.
Unless they pull off yet another breakthrough, I don’t think they have much of a great future
Investors OTOH...
Suleyman’s team has also been testing alternatives from companies like xAI, DeepSeek, and Meta
Also a bit hyperbolic. I'm sure there are good reasons Microsoft would want to build it's own products on top of their own models and have more fine control of things. That doesn't mean they are plotting a future where they do nothing at all with OpenAI.
Watch.
Nadella will not steer this correctly