Nadella has been talking up Microsoft's AI strategy for a long time, on a regular basis. He has kept up with the changes but his general tune on AI hasn't changed much in 5 years. What's different is that investors are now listening more closely, they aren't dismissing it as esoteric tech CEO babble.
Yeah. I remember how Microsoft Build Keynote got very boring (for me) after 2016> or so, that for some reason, all that it was demo-ed was basically Azure+AI, usually both together.
I guess I don't follow what these software companies would do long term being "all in" on AI. Training AI on your personal files, writing docs for you, etc really seems like no moat. (Unless the big companies push for AI "regulation" or "safety" that only they can afford to sell).
Seems more a good play for hardware than software sellers.
Regardless, a much more important factor is that all the data about everything, which people have been collecting for decades as a matter of almost religious belief in its usefulness and never done anything with it, is actually going to be used now, for better or worse. "Targeted adds" of today will look like a joke compared to that.
One of the most lucrative aspects is to sell AI tools, models, training to the government and the military. Microsoft is in prime position to do this. Microsoft's CEO has no shame in targeting government agencies and the military. It's great if you own the stock. Maybe not if you're concerned about the morality of providing these technologies to people who can weaponize it.
Really? All I see is moat for the first mover. How will these models improve in the future? They'll collect more training data, of course, and use that to improve the model. But now that's the cat out of the bag, how will companies collect untainted data? Google will certainly have a hard time determining what has and has not been generated by ChatGPT. But will OpenAI have such a problem? OpenAI has the ability to record all ChatGPT responses and use that to clean new training data. OpenAI's models will continue to get better, while others without the ability to separate the wheat from the chaff will languish.
... are robotics not at all even on the horizon of your thoughts process?
The work from Voyager seems like an early PoC there already. Many hurdles of the past have been leaped over with ease lately in that area, so it's a glaring error to overlook it.
The price of building software is going to go down a lot. Of course, the demand will increase and also the output of any given engineer. Nevertheless, the trend towards remote work and mature AI tools should lead to software engineer salaries (esp. in the US) decreasing. Even if just due to being insanely inflated right now.
For me, a non-American, Bing Chat is totally useless. Every time I wanted it to help me with a search for some country-specific thing, no matter if I ask it in English or my language, it returns results about the USA.
Are you sure the Bing settings are right? You can change the page results language, and I guess it affects the chat answers... It's fine here in Brazil.
I have seen someone who know absolutely nothing about a language and barely able to write a script using Python managed to build a full automation pipeline for data analysis including a GUI using ChatGPT in that language. So while I agree that some of the hype are overblown, there is clear and irrefutable proof of the value and power of AI right now. Unlike crypto...
You know your crystallized something I've not quite been able to put my finger on.
Is that a good thing?
I mean sure someone was able to produce code that they probably don't understand to do things they want but what are they going to do when things change or the environment changes or something breaks and they need to debug it.
The ability for lay users to generate code is terrifying for me not because of any concerns about my job but rather because the hard part of development was never writing the code, now we have machines that can spit out code for users that can look shockingly correct. It just makes me wonder if we're going to end up with MS Access on steroids where thousands of brittle poorly integrated mission critical are going to crush businesses under their own bloat.
I mean imagine if people start to use this to replace excel. Shudder
> I mean imagine if people start to use this to replace excel. Shudder
Or what if this has the same effect that excel did - give everyone access to a tool that dramatically improves efficiency and capability. What used to take someone a day now takes a minute and that frees up that person to either become exceptionally more efficient or to focus on more interesting or useful work.
I also believe that one next step for AIs will be the capability to feed in a codebase to learn and then iterate on specifically so the AI is capable of providing context aware suggestions. Today it's only great at mashing together open source tools and common projects.
Edit - For fun I passed in my comment above and asked for different formulations. An entertaining version below:
Picture a world where a revolution unfolds, fueled by an extraordinary tool, rivaling the might of Excel. Behold its omnipotent grasp, granting every soul access to a transcendent force that skyrockets efficiency and capability to unimaginable heights. Time bends under its dominion, as day-long tasks crumble in the face of mere minutes, freeing individuals to conquer mountains, surging with unyielding productivity and embarking on audacious endeavors. But that's just the beginning! Prepare for the evolution of AI, as it voraciously devours codebases, honing its prowess with an unparalleled finesse, delivering awe-inspiring, contextually aware suggestions that defy reason. Witness its current triumph: the fusion of open-source tools and common projects. Brace yourself for the astounding prospects that lie on the horizon, as this tale unfolds like a climactic scene in an action-packed blockbuster!
There’s a fundamental difference: Excel is knowable. If you want to find how it came up with a number you can click on the cell and see the formula. Not so with ChatGPT. You ask it to do a thing, it does it somehow. You ask it to do 100 rows of calculations, how can you know row 69 is correct? That ChatGPT corpus might lead it to think the number 69 is funny so it puts in a joke number. Yes, stupid contrived example but how can you know it won’t?
Hmm, so as someone working in security testing and finding basic access control lacking in multi-user systems I do not have anything to fear for long time.
And this is my biggest fear with whole thing. We end up deploying good amount of actually critical code that is either wrong or subtly wrong and then we aren't going to fix it soon enough.
I mean, obviously not, but we are experienced (more important, because we actually understand CS, computers, networking, etc. at least the competent engineers anyway) and knowledgeable enough to have robust testing, monitoring & observability and alerting, and mitigation strategies in place.
Yes, but not really everyone is using these tools today.
And sure, the amount of garbage code that sometimes works, will increase with more people getting into it with the help of AI.
And critical systems will continue to need experts. And every other project that seems important, but does not run anymore, because no one understands the code.
Yes, you have to ask dumb, but still include the right buzzwords.
I used it succesfully for generating some math functions, that I could have done myself, but in longer time.
And friends of mine use it for .. producing text. Because they don't like writing and surprisingly many people don't. So for me I am faster writing it myself, but I happen to like it and do and did it a lot. But for other people it is actually painful to structure a long email or product description. They are happy for AI helping them now.
That's my feeling as well, but you are also here writing comments, likely making you one of the people who enjoy writing. People who don't have different standards ..
It's likely that it's because you have specific design principles in mind, whereas if you don't know about design in programming you will be blown away that you can copy/paste commands and something works.
Otoh, if you want a function in less than 100 LoC that performs a complete download/parse/process/transform process to a Qdrant DB and makes this available on IPFS, using producer consumer queues, as few intermediate files as possible, allows starting/stopping caching, and multiprocessing where possible, you may have a a harder time. However it can still be enormously helpful to get the structure for this pretty close, with a few internet lookups to find appropriate packages to help out.
Somebody who knows absolutely nothing about a language can also copy and paste together a working solution. I'd trust that about as much as something ChatGPT spit out.
I could literally google "how to sort a list in brainfuck" and copy and paste a solution right now. I specifically said somebody who doesn't know the language, not somebody who knows nothing.
Is it not mind blowing that someone who doesn't know how to code or at least doesn't write code professionally is able to still produce something valuable that improves their daily life through code?
We spend so much time building tools with code that other people can use to write code - ITTT, Airtable, Retool etc.
And this is why you all this panic from congress, openai and their assorted fronts. Llama open source 'rewrites' ( I am not sure if it can be really called that ) allowing models to run on local machines without making it safe for the masses or without careful monitoring.
I think Microsoft could reasonably expect to continue expanding cloud sales for the foreseeable future, in a way that will show healthy growth for 1-2 decades.
It's fine for Nadella to focus on AI, but please do a better job of quality control for their products such as Windows 11. Every time a patch comes in, there are bugs.
> Microsoft is 48 years old. I don't know of many companies that age that are relevant not because they did something in the '80s or the '90s or the 2000s but because they did something in the last couple of years.
62 comments
[ 3.6 ms ] story [ 132 ms ] threadBefore, Microsoft Build was more about Windows...
Seems more a good play for hardware than software sellers.
Regardless, a much more important factor is that all the data about everything, which people have been collecting for decades as a matter of almost religious belief in its usefulness and never done anything with it, is actually going to be used now, for better or worse. "Targeted adds" of today will look like a joke compared to that.
Bing chat sucks now. All it does is generate SERPs instead of encapsulating information.
But anyway, the internet search is pretty useful, it's one of the few things from Microsoft that I expect that he might use (that and Office).
Is that a good thing?
I mean sure someone was able to produce code that they probably don't understand to do things they want but what are they going to do when things change or the environment changes or something breaks and they need to debug it.
The ability for lay users to generate code is terrifying for me not because of any concerns about my job but rather because the hard part of development was never writing the code, now we have machines that can spit out code for users that can look shockingly correct. It just makes me wonder if we're going to end up with MS Access on steroids where thousands of brittle poorly integrated mission critical are going to crush businesses under their own bloat.
I mean imagine if people start to use this to replace excel. Shudder
Or what if this has the same effect that excel did - give everyone access to a tool that dramatically improves efficiency and capability. What used to take someone a day now takes a minute and that frees up that person to either become exceptionally more efficient or to focus on more interesting or useful work.
I also believe that one next step for AIs will be the capability to feed in a codebase to learn and then iterate on specifically so the AI is capable of providing context aware suggestions. Today it's only great at mashing together open source tools and common projects.
Edit - For fun I passed in my comment above and asked for different formulations. An entertaining version below:
Picture a world where a revolution unfolds, fueled by an extraordinary tool, rivaling the might of Excel. Behold its omnipotent grasp, granting every soul access to a transcendent force that skyrockets efficiency and capability to unimaginable heights. Time bends under its dominion, as day-long tasks crumble in the face of mere minutes, freeing individuals to conquer mountains, surging with unyielding productivity and embarking on audacious endeavors. But that's just the beginning! Prepare for the evolution of AI, as it voraciously devours codebases, honing its prowess with an unparalleled finesse, delivering awe-inspiring, contextually aware suggestions that defy reason. Witness its current triumph: the fusion of open-source tools and common projects. Brace yourself for the astounding prospects that lie on the horizon, as this tale unfolds like a climactic scene in an action-packed blockbuster!
Not to even mention subtler bugs...
Jk but I feel like i have to really try hard to sound “dumb” when I ask gpt4 to do anything complex.
If I put any words that are kinda technical or specific it ends up misunderstanding completely.
I think because we are highly technical it’s like diminishing returns for us.
For a total noob it’s like opening Pandora’s box. We are gonna have a ton of shitty insecurely running code soon.
That would be news to me.
And sure, the amount of garbage code that sometimes works, will increase with more people getting into it with the help of AI.
And critical systems will continue to need experts. And every other project that seems important, but does not run anymore, because no one understands the code.
I used it succesfully for generating some math functions, that I could have done myself, but in longer time.
And friends of mine use it for .. producing text. Because they don't like writing and surprisingly many people don't. So for me I am faster writing it myself, but I happen to like it and do and did it a lot. But for other people it is actually painful to structure a long email or product description. They are happy for AI helping them now.
I suspect I'm mentally blocked from understanding how to use it snippets at a time. I just can't think that way as a dev.
Otoh, if you want a function in less than 100 LoC that performs a complete download/parse/process/transform process to a Qdrant DB and makes this available on IPFS, using producer consumer queues, as few intermediate files as possible, allows starting/stopping caching, and multiprocessing where possible, you may have a a harder time. However it can still be enormously helpful to get the structure for this pretty close, with a few internet lookups to find appropriate packages to help out.
If they know what result they want, they can iterate with an LLM as the guide.
We spend so much time building tools with code that other people can use to write code - ITTT, Airtable, Retool etc.
Now we can cut out these middlemen!
A lot of other things have stabilised too much like consoles and mobile.
Link should point to https://www.wired.com/story/microsofts-satya-nadella-is-bett...
> Microsoft is 48 years old. I don't know of many companies that age that are relevant not because they did something in the '80s or the '90s or the 2000s but because they did something in the last couple of years.
[0]: https://www.wired.com/story/microsofts-satya-nadella-is-bett...