I’m 20 minutes into the video and it does seem mostly basic and agreeable.
Two arguments from Ng that really stuck out that is really tripping my skepticism alarm are:
1) He mentions how fast prototyping has begun because generating a simple app has become easier with AI. This, to me, has always been quick and never the bottleneck for any company I’ve been at, including startups. Validating an idea was simple enough via wireframing. I can maybe see it for selling an idea where you need some amount of fidelity yo impress potential investors… but I would hope places like YC can see the tech behind the idea without seeing the tech itself. Or at least can ignore low fidelity if a prototype shows the meat of the product.
2) Ng talks about how everyone in his company codes, from the front desk to the executives. The “everyone should code” idea has been done and shown to fail for the past 15 years. In fact I’ve seen it be more damaging than helpful because it gave people false confidence that they could tell engineers how to do their job rather than a more empathetic understanding.
At my company everybody codes, including PMs and business people. It can definitely be damaging done in the long run without any supervision from an actual programmer. This is why we assign an engineer to review every PR of a vibe coded project and they don’t really need all of the context to detect bs approaches that will surely fail.
About prototyping - its much faster and i dont know how anyone can argue this. PMs can get a full blown prototype for an MVP working in a day with AI assistance. Sure - they will be thrown in the trash after the demo, but they carry out their purpose of proving a concept. The code is janky but it works for its purpose.
Good lord I think I'd rather eat a shotgun than be forced to review a billion garbage PRs made by PMs and other non-technical colleagues. It's bad enough reviewing PRs backenders writing frontend features badly (and vice versa), I cannot even imagine the pits of hell this crap is like.
> This is why we assign an engineer to review every PR of a vibe coded project and they don’t really need all of the context to detect bs approaches that will surely fail.
I see this trend in many companies as well, just curious, how do you make sure engineering time is not wasted reviewing so many PRs? Because, some of them will be good, you only need couple of your bets to take off, some definitely bad
On point 1, it's worse than that. Adding detail and veracity to a prototype is well known to bring negative value.
Prototypes must be exactly a sketchy as the ideas they represent, otherwise they mislead people into thinking the software is built and your ideas can't be changed.
I’ve always said this as well, having done lots and lots of early stage building and prototyping, and suffering plenty of proto-duction foibles, however my view has shifted on this a lot in the last year or so.
With current models I’m able to throw together fully working web app prototypes so quickly and iterate often-sweeping UI and architectural changes so readily that I’m finding it has changed my whole workflow. The idea of trying to keep things low-fidelity at the start is predicated on the understanding that changes later in the process are much more difficult or expensive, which I think is increasingly no longer the case in many circumstances. Having a completely working prototype and then totally changing how it works in just a few sentences is really quite something.
The key to sustainability in this pattern, in my opinion, is not letting the AI dictate project structure or get too far ahead of your own understanding/oversight of the general architecture. That’s a balancing act to be sure, since purely vibe-coding is awfully tempting, but it’s still far too easy to wind up with a big ball of wax that neither human nor AI can further improve.
My two takeaways is you build
1) Having a precise vision of what you want to achieve
2) Being able to control / steer AI towards that vision
Teams that can do both of these things, especially #1 will move much faster. Even if they are wrong its better than vague ideas that get applause but not customers
Yes this! The observation that being specific versus general in the problems you want to solve is a better start-up plan is true for all startups ever, not just ones that use LLMs to solve them. Anecdotal/personal startup experiences support this strongly and I read enough on here to know that I am not alone…
What's the balance between being specific in a way that's positive and allows you to solve good problems, and not getting pigeonhold and not being able to pivot? I wonder if companies who pivot are the norm or if you just here of the most popular cases.
Not sure why this has drawn silence and attacks - whence the animus to Ng? His high-level assessments seem accurate, he's a reasonable champion of AI, and he speaks credibly based on advising many companies. What am I missing? (He does fall on the side of open models (as input factors): is that the threat?)
He argues that landscape is changing (at least quarterly), and that services are (best) replaceable (often week-to-week) because models change, but that orchestration is harder to replace, and that there are relatively few orchestration platforms.
So: what platforms are available? Are there other HN posts that assess the current state of AI orchestration?
(What's the AI-orchestration acronym? not PAAS but AIOPAAS? AOP? (since aspect-oriented programming is history))
I'm guessing because this is basically an AI for Dummies overview, while half of HN is deep in the weeds with AI already. Nothing wrong with the talk! Except his focus on "do everything" agents already feels a bit stale as the move seems to be going in the direction of limited agents with a much stronger focus on orchestration of tools and context.
> I'm guessing because this is basically an AI for Dummies
I second this, for the silence at least, I listened to the talk because it was Andrew Ng and it is good or at least fun to listen to talks by famous people, but I did not walk away with any new key insights, which is fine, most talks are not that.
And he’s been doing it forever and all from the original idea that he could offer a Stanford education on ai for free on the Internet thus he created coursera. The dude is cool.
Yea haha the chinese-to-english gets confusing, because it's not a 1:1, it's an N:1 thing, for the number different Chinese languages, different tones, and semi-malicious US immigration agents who botched the shit out of people's names in the late 19th and early 20th century.
Wu and Ng in Mandarin and Cantonese may be the same character. But Wu the common surname and Wu for some other thing (e.g. that mountain) may be different characters entirely.
It gets even more confusing when you throw a third Chinese language in, say Taishanese:
Wu = Ng (typically) for Mandarin and Cantonese et al. But if it's someone who went to America earlier, suddenly it's Woo. But even though they're both yue Chinese languages, Woo != Woo in Cantonese and Taishanese. For that name, it's Hu (Mandarin) = Wu / Wuh (Cantonese) = Woo (Taishanese, in America). Sometimes. Lol. Sometimes not.
And between that and the rap group there’s this important movie:
Shaolin and Wu Tang (1983)
> The film is about the rivalry between the Shaolin (East Asian Mahayana) and Wu-Tang (Taoist Religion) martial arts schools. […]
> East Coast hip-hop group Wu-Tang Clan has cited the film as an early inspiration. The film is one of Wu-Tang Clan founder RZA's favorite films of all time. Founders RZA and Ol' Dirty Bastard first saw the film in 1992 in a grindhouse cinema on Manhattan's 42nd Street and would found the group shortly after with GZA. The group would release its debut album Enter the Wu-Tang (36 Chambers), featuring samples from the film's English dub; the album's namesake is an amalgamation of Enter the Dragon (1973), Shaolin and Wu Tang, and The 36th Chamber of Shaolin (1978).
I couldn't tell you, but what I can contribute to that discussion is that orchestration of AI in its current form would focus on one of two approaches: consistent output despite the non-deterministic state of LLMs, or consistent inputs that leans into the non-deterministic state of LLMs. The problem with the former (output) is that you cannot guarantee the output of an AI on a consistent basis, so a lot of the "orchestration" of outputs is largely just brute-forcing tokens until you get an answer within that acceptable range; think the glut of recent "Show HN" stuff where folks built a slop-app by having agents bang rocks together until the code worked.
On the input side of things, orchestration is less about AI itself and more about ensuring your data and tooling is consistently and predictably accessible to the AI such that the output is similarly predictable or consistent. If you ask an AI what 2+2 is a hundred different ways, you increase the likelihood of hallucinations; on the other hand, ensuring the agent/bot gets the same prompt with the same data formats and same desired outputs every single time makes it more likely that it'll stay on task and not make shit up.
My engagement with AI has been more of the input-side, since that's scalable with existing tooling and skillsets in the marketplace instead of the output side, which requires niche expertise in deep learning, machine learning, model training and fine-tuning, etc. In other words, one set of skills is cheaper and more plentiful while also having impacts throughout the organization (because everyone benefits from consistent processes and clean datasets), while the other is incredibly expensive and hard to come by with minimal impacts elsewhere unless a profound revolution is achieved.
One thing to note is that Dr. Ng gives the game away at the Q&A portion fairly early on: "In the future, the people who are the most powerful are the people who can make computers do exactly what you want it to do." In that context, the current AI slop is antithetical to what he's pitching. Sure, AI can improve speed on execution, prototyping, and rote processes, but the real power remains in the hands of those who can build with precision instead of brute-force. As we continue to hit barriers in the physical capabilities of modern hardware and wrestle with the effects of climate change and/or poor energy policies, efficiency and precision will gradually become more important than speed - at least that's my thinking.
Really valid points. I agree with the bits about “expertise in getting the computer to do what you want” being the way of the future, but he also raises really valid points about people having strong domain knowledge (a la his colleague with extensive art history knowledge being better at midjourney than him) after saying it’s okay to tell people to just let the LLM write code for you and learn to code that way. I am having a hard time with the contradictions, maybe it’s me. Not meaning to rag on Dr. Ng, just further the conversation. (Which is super interesting to me.)
EDIT: rereading and realizing I think what resonates most is we are in agreement about the antithetical aspects of the talk. I think this is the crux of the issue.
> The problem with the former (output) is that you cannot guarantee the output of an AI on a consistent basis
Do you mean you cannot guarantee the result based on a task request with a random query? Or something else? I was under the impression that LLMs are very deterministic if you provide a fixed seed for the samplers, fixed model weights, and fixed context. In cloud providers you can't guarantee this because of how they implement this (batching unrelated requests together and doing math). Now you can't guarantee the quality of the result from that and changing the seed or context can result in drastically different quality. But maybe you really mean non-deterministic but I'm curious where this non-determinism would come from.
> I was under the impression that LLMs are very deterministic if you provide a fixed seed for the samplers, fixed model weights, and fixed context.
That's all input-side, though. On the output side, you can essentially give an LLM anxiety by asking the exact same question in different ways, and the machine doesn't understand anymore that you're asking the exact same question.
For instance, take one of these fancy "reasoning" models and ask it variations on 2+2. Try two plus two, 2 plus two, deux plus 2, TwO pLuS 2, etc, and observe its "reasoning" outputs to see the knots it ties itself up in trying to understand why you keep asking the same calculation over and over again. Running an older DeepSeek model locally, the "reasoning" portion continued growing in time and tokens as it struggled to provide context that didn't exist to a simple problem that older/pre-AI models wouldn't bat an eye at and spit out "4".
Trying to wrangle consistent, reproducible outputs from LLMs without guaranteeing consistent inputs is a fool's errand.
Ok yes. I call that robustness of the model as opposed to determinism which to me implies different properties. And yes, I too have been frustrated by the lack of robustness of models to minor variations in input or even using a different seed for the same input.
Pointing out that LLMs are deterministic as long as you lock down everything, is like saying an extra bouncy ball doesn’t bounce if you leave it on flat surface, reduce the temperature to absolute zero, and make sure the surface and the ball are at rest before starting the experiment.
It’s true but irrelevant.
One of the GP’s main points was that even the simplest questions can lead to hundreds of different contexts; they probably already know that you could get different outcomes if you could instead have a fixed context.
The platforms I've seen live on top of kubernetes so I'm afraid it is possible. nvidia-docker, all the cuda libraries and drivers, nccl, vllm,... Large scale distributed training and inference are complicated beasties and the orchestration for them is too.
AOP always felt like a hack. I used it with C++ early on, and it was a preprocessor inserting ("weaving") aspects in the function entries/exits. Mostly was useful for logging. But that can be somewhat emulated using C++ constructors/destructors.
Maybe it can be also useful for DbC (Design-by-Contract) when sets of functions/methods have common pre/post-conditions and/or invariants.
This talk is deceptively simple. The most sage advice that founders routinely forget is what concrete idea are you going to implement and why do you think it will work? There has be a way to invalidate your idea and as a corollary you must have the focus to collect the data and properly invalidate it.
I have had reservation about Ng from a lot of his past hype, but I thought this talk was extremely practical and tactical. I recommend watching it before passing judgement.
not a single word about overwhelming replacement of humans with AI. nothing about countless jobs lost. nothing about ever increasing competition and rat-race. (speaking of software, but applies to all industries). his rose-glasses view is somewhere in between optimism-in-denial to straight-up lunacy. if this is the leader(s) we have been following, this should be a wake up call.
A good chunk of Ng's work these days seems to be around AI Fund [0] which he explicitly mentioned in the video, in the first 5 seconds, involves co-founding these startups and being in the weeds with the initial development.
Additionally, he does engage pretty closely with the teams behind the content of his deeplearning.ai lectures and does make sure he has a deep understanding of the products these companies are highlighting.
He certainly is a businessman, but that doesn't exlcudethe possibility that he remains highly knowledgeable about this space.
Except they aren't pay to play unless you consider doing the work for the course the "payment". There's certainly an exchange since there is a lot of work involved, but DLAI provides a team to help design, structure and polish the course and then the team creating the course does the majority of the work creating the content, but there's no financial exchange.
The DLAI team is also pretty good about ensuring the content covers a topic not a product in general.
The content is a repackage of previously existing, publicly available notebooks, docs, YouTube videos. I wouldnt be surprised if the repackaging was done by AI.
Ng built baidu's AI department and began their start in various sectors with actual AI system design, so yes, he isn't a failed startup entrepreneur like any vibe startup maker who already wants to stop and give advice.
Maybe you can help me hire a vibe coder with 10 years experience?
Right.. He's just a giant, not a midget with a step ladder.
But I do question why anyone who played a significant role in the foundation of the current AI generation would teach an obvious new Zuckerberg generation who will apparently think they are the start of everything if they get a style working in the prompt.
If not for 3 people in 2012, I find it highly unlikely a venture like OpenAI could have occurred and without Ng in particular I wouldn't be surprised if the field would have been missing a few technical pieces as well as the hire-able engineers.
He sold courses (great ones!) long before there was AI-gold rush. He's one of the OG players in online education and I think he deserves praise, not blame for that.
I think this is an interesting question, and I’d like to genuinely attempt an answer.
I essentially think this is because people prefer to optimize what they can measure.
It is hard to measure the quality of work. People have subjective opinions, the size of opportunities can be different, etc, making quality hard to pin down. It is much easier to measure the time required for each iteration on a concept. Additionally, I think it is generally believed that a project with more iterations tends to have higher quality than a project with less, even putting aside the concern about measuring quality itself. Therefore, we put aside the discussion of quality (which we’d really like to improve), and instead make the claim that we can actually measure (time to do something), with the strong implication that this _also_ will tend to increase quality.
Energy consumption and data protection were a thing and then came AI and all of a sudden it doesn’t matter anymore.
Between all the good things people create with AI I see a lot more useless or even harmful things.
Scams and fake news get better and harder to distinguish to a point where reality doesn’t matter anymore.
I think quality takes time and refinement which is not something that LLMs have solved very well today. They are very okay at it, except for very specific targeted refinements (Grammerly, SQL editors).
However, they are excellent at building from 0->1, and the video is suggesting that this is perfect for startups. In the context of startups, faster is better.
I’m 20 minutes into the video and it does seem mostly basic and agreeable.
Two arguments from Ng that really stuck out that is really tripping my skepticism alarm are:
1) He mentions how fast prototyping has begun because generating a simple app has become easier with AI. This, to me, has always been quick and never the bottleneck for any company I’ve been at, including startups. Validating an idea was simple enough via wireframing. I can maybe see it for selling an idea where you need some amount of fidelity yo impress potential investors… but I would hope places like YC can see the tech behind the idea without seeing the tech itself. Or at least can ignore low fidelity if a prototype shows the meat of the product.
2) Ng talks about how everyone in his company codes, from the front desk to the executives. The “everyone should code” idea has been done and shown to fail for the past 15 years. In fact I’ve seen it be more damaging than helpful because it gave people false confidence that they could tell engineers how to do their job rather than a more empathetic understanding.
At my company everybody codes, including PMs and business people. It can definitely be damaging done in the long run without any supervision from an actual programmer. This is why we assign an engineer to review every PR of a vibe coded project and they don’t really need all of the context to detect bs approaches that will surely fail.
About prototyping - its much faster and i dont know how anyone can argue this. PMs can get a full blown prototype for an MVP working in a day with AI assistance. Sure - they will be thrown in the trash after the demo, but they carry out their purpose of proving a concept. The code is janky but it works for its purpose.
Good lord I think I'd rather eat a shotgun than be forced to review a billion garbage PRs made by PMs and other non-technical colleagues. It's bad enough reviewing PRs backenders writing frontend features badly (and vice versa), I cannot even imagine the pits of hell this crap is like.
> This is why we assign an engineer to review every PR of a vibe coded project and they don’t really need all of the context to detect bs approaches that will surely fail.
I see this trend in many companies as well, just curious, how do you make sure engineering time is not wasted reviewing so many PRs? Because, some of them will be good, you only need couple of your bets to take off, some definitely bad
On point 1, it's worse than that. Adding detail and veracity to a prototype is well known to bring negative value.
Prototypes must be exactly a sketchy as the ideas they represent, otherwise they mislead people into thinking the software is built and your ideas can't be changed.
I’ve always said this as well, having done lots and lots of early stage building and prototyping, and suffering plenty of proto-duction foibles, however my view has shifted on this a lot in the last year or so.
With current models I’m able to throw together fully working web app prototypes so quickly and iterate often-sweeping UI and architectural changes so readily that I’m finding it has changed my whole workflow. The idea of trying to keep things low-fidelity at the start is predicated on the understanding that changes later in the process are much more difficult or expensive, which I think is increasingly no longer the case in many circumstances. Having a completely working prototype and then totally changing how it works in just a few sentences is really quite something.
The key to sustainability in this pattern, in my opinion, is not letting the AI dictate project structure or get too far ahead of your own understanding/oversight of the general architecture. That’s a balancing act to be sure, since purely vibe-coding is awfully tempting, but it’s still far too easy to wind up with a big ball of wax that neither human nor AI can further improve.
even prototyping hasn't become "fast" because you cannot purely vibecode even a prototype.
My two takeaways is you build 1) Having a precise vision of what you want to achieve 2) Being able to control / steer AI towards that vision
Teams that can do both of these things, especially #1 will move much faster. Even if they are wrong its better than vague ideas that get applause but not customers
Yes this! The observation that being specific versus general in the problems you want to solve is a better start-up plan is true for all startups ever, not just ones that use LLMs to solve them. Anecdotal/personal startup experiences support this strongly and I read enough on here to know that I am not alone…
What's the balance between being specific in a way that's positive and allows you to solve good problems, and not getting pigeonhold and not being able to pivot? I wonder if companies who pivot are the norm or if you just here of the most popular cases.
Not sure why this has drawn silence and attacks - whence the animus to Ng? His high-level assessments seem accurate, he's a reasonable champion of AI, and he speaks credibly based on advising many companies. What am I missing? (He does fall on the side of open models (as input factors): is that the threat?)
He argues that landscape is changing (at least quarterly), and that services are (best) replaceable (often week-to-week) because models change, but that orchestration is harder to replace, and that there are relatively few orchestration platforms.
So: what platforms are available? Are there other HN posts that assess the current state of AI orchestration?
(What's the AI-orchestration acronym? not PAAS but AIOPAAS? AOP? (since aspect-oriented programming is history))
I'm guessing because this is basically an AI for Dummies overview, while half of HN is deep in the weeds with AI already. Nothing wrong with the talk! Except his focus on "do everything" agents already feels a bit stale as the move seems to be going in the direction of limited agents with a much stronger focus on orchestration of tools and context.
> I'm guessing because this is basically an AI for Dummies
I second this, for the silence at least, I listened to the talk because it was Andrew Ng and it is good or at least fun to listen to talks by famous people, but I did not walk away with any new key insights, which is fine, most talks are not that.
From the recent threads, it feels like the other half is totally, willfully ignorant. Hence the responses.
As someone who is part of that other half, I agree.
> deep in the weeds with AI already
I doubt even 10% have written a custom MCP tool... and probably some who don't even know what that means
I like Andrew Ng. He's like the Mister Rogers of AI. I always listen when he has something to say.
And he’s been doing it forever and all from the original idea that he could offer a Stanford education on ai for free on the Internet thus he created coursera. The dude is cool.
Is he affiliated with nghttp?
No?
ng*, ng-*, or *-ng is typically "Next Generation" in software nomenclature. Or, star trek (TNG). Alternatively, "ng-" is also from angular-js.
Ng in Andrew Ng is just his name, like Wu in Chinese.
Wu from Wu-Tang?
No, Wu-Tang ultimately derives from the Wudang Mountains, with the corresponding Cantonese being Moudong https://en.wiktionary.org/wiki/%E6%AD%A6%E7%95%B6%E5%B1%B1
Yea haha the chinese-to-english gets confusing, because it's not a 1:1, it's an N:1 thing, for the number different Chinese languages, different tones, and semi-malicious US immigration agents who botched the shit out of people's names in the late 19th and early 20th century.
Wu and Ng in Mandarin and Cantonese may be the same character. But Wu the common surname and Wu for some other thing (e.g. that mountain) may be different characters entirely.
It gets even more confusing when you throw a third Chinese language in, say Taishanese:
Wu = Ng (typically) for Mandarin and Cantonese et al. But if it's someone who went to America earlier, suddenly it's Woo. But even though they're both yue Chinese languages, Woo != Woo in Cantonese and Taishanese. For that name, it's Hu (Mandarin) = Wu / Wuh (Cantonese) = Woo (Taishanese, in America). Sometimes. Lol. Sometimes not.
Similarly, Mei = Mai = Moy
And between that and the rap group there’s this important movie:
Shaolin and Wu Tang (1983)
> The film is about the rivalry between the Shaolin (East Asian Mahayana) and Wu-Tang (Taoist Religion) martial arts schools. […]
> East Coast hip-hop group Wu-Tang Clan has cited the film as an early inspiration. The film is one of Wu-Tang Clan founder RZA's favorite films of all time. Founders RZA and Ol' Dirty Bastard first saw the film in 1992 in a grindhouse cinema on Manhattan's 42nd Street and would found the group shortly after with GZA. The group would release its debut album Enter the Wu-Tang (36 Chambers), featuring samples from the film's English dub; the album's namesake is an amalgamation of Enter the Dragon (1973), Shaolin and Wu Tang, and The 36th Chamber of Shaolin (1978).
https://en.wikipedia.org/wiki/Shaolin_and_Wu_Tang
> So: what platforms are available?
I couldn't tell you, but what I can contribute to that discussion is that orchestration of AI in its current form would focus on one of two approaches: consistent output despite the non-deterministic state of LLMs, or consistent inputs that leans into the non-deterministic state of LLMs. The problem with the former (output) is that you cannot guarantee the output of an AI on a consistent basis, so a lot of the "orchestration" of outputs is largely just brute-forcing tokens until you get an answer within that acceptable range; think the glut of recent "Show HN" stuff where folks built a slop-app by having agents bang rocks together until the code worked.
On the input side of things, orchestration is less about AI itself and more about ensuring your data and tooling is consistently and predictably accessible to the AI such that the output is similarly predictable or consistent. If you ask an AI what 2+2 is a hundred different ways, you increase the likelihood of hallucinations; on the other hand, ensuring the agent/bot gets the same prompt with the same data formats and same desired outputs every single time makes it more likely that it'll stay on task and not make shit up.
My engagement with AI has been more of the input-side, since that's scalable with existing tooling and skillsets in the marketplace instead of the output side, which requires niche expertise in deep learning, machine learning, model training and fine-tuning, etc. In other words, one set of skills is cheaper and more plentiful while also having impacts throughout the organization (because everyone benefits from consistent processes and clean datasets), while the other is incredibly expensive and hard to come by with minimal impacts elsewhere unless a profound revolution is achieved.
One thing to note is that Dr. Ng gives the game away at the Q&A portion fairly early on: "In the future, the people who are the most powerful are the people who can make computers do exactly what you want it to do." In that context, the current AI slop is antithetical to what he's pitching. Sure, AI can improve speed on execution, prototyping, and rote processes, but the real power remains in the hands of those who can build with precision instead of brute-force. As we continue to hit barriers in the physical capabilities of modern hardware and wrestle with the effects of climate change and/or poor energy policies, efficiency and precision will gradually become more important than speed - at least that's my thinking.
Really valid points. I agree with the bits about “expertise in getting the computer to do what you want” being the way of the future, but he also raises really valid points about people having strong domain knowledge (a la his colleague with extensive art history knowledge being better at midjourney than him) after saying it’s okay to tell people to just let the LLM write code for you and learn to code that way. I am having a hard time with the contradictions, maybe it’s me. Not meaning to rag on Dr. Ng, just further the conversation. (Which is super interesting to me.)
EDIT: rereading and realizing I think what resonates most is we are in agreement about the antithetical aspects of the talk. I think this is the crux of the issue.
> The problem with the former (output) is that you cannot guarantee the output of an AI on a consistent basis
Do you mean you cannot guarantee the result based on a task request with a random query? Or something else? I was under the impression that LLMs are very deterministic if you provide a fixed seed for the samplers, fixed model weights, and fixed context. In cloud providers you can't guarantee this because of how they implement this (batching unrelated requests together and doing math). Now you can't guarantee the quality of the result from that and changing the seed or context can result in drastically different quality. But maybe you really mean non-deterministic but I'm curious where this non-determinism would come from.
> I was under the impression that LLMs are very deterministic if you provide a fixed seed for the samplers, fixed model weights, and fixed context.
That's all input-side, though. On the output side, you can essentially give an LLM anxiety by asking the exact same question in different ways, and the machine doesn't understand anymore that you're asking the exact same question.
For instance, take one of these fancy "reasoning" models and ask it variations on 2+2. Try two plus two, 2 plus two, deux plus 2, TwO pLuS 2, etc, and observe its "reasoning" outputs to see the knots it ties itself up in trying to understand why you keep asking the same calculation over and over again. Running an older DeepSeek model locally, the "reasoning" portion continued growing in time and tokens as it struggled to provide context that didn't exist to a simple problem that older/pre-AI models wouldn't bat an eye at and spit out "4".
Trying to wrangle consistent, reproducible outputs from LLMs without guaranteeing consistent inputs is a fool's errand.
Ok yes. I call that robustness of the model as opposed to determinism which to me implies different properties. And yes, I too have been frustrated by the lack of robustness of models to minor variations in input or even using a different seed for the same input.
Pointing out that LLMs are deterministic as long as you lock down everything, is like saying an extra bouncy ball doesn’t bounce if you leave it on flat surface, reduce the temperature to absolute zero, and make sure the surface and the ball are at rest before starting the experiment.
It’s true but irrelevant.
One of the GP’s main points was that even the simplest questions can lead to hundreds of different contexts; they probably already know that you could get different outcomes if you could instead have a fixed context.
This is great thinking, thank you for writing this.
We've defined agents. Let's now define orchestration.
Bold claim. I am not convinced anyone's done a good job defining agents and if they did 99% of the population has a different interpretation.
Okay. We've tried to define agents. Now let's try to define orchestration.
And make it more complicated than K8s
Not possible
The platforms I've seen live on top of kubernetes so I'm afraid it is possible. nvidia-docker, all the cuda libraries and drivers, nccl, vllm,... Large scale distributed training and inference are complicated beasties and the orchestration for them is too.
> AOP? (since aspect-oriented programming is history)
AOP is very much alive, people that do AOP have just forgotten what the name is, and many have simply reinvented it poorly.
AOP always felt like a hack. I used it with C++ early on, and it was a preprocessor inserting ("weaving") aspects in the function entries/exits. Mostly was useful for logging. But that can be somewhat emulated using C++ constructors/destructors.
Maybe it can be also useful for DbC (Design-by-Contract) when sets of functions/methods have common pre/post-conditions and/or invariants.
https://en.wikipedia.org/wiki/Aspect-oriented_programming#Cr...
Also very much alive and called that in the Java/Spring ecosystem
No need to add AI to the name, especially if it works. PaaS and IaaS are sufficient.
[flagged]
This talk is deceptively simple. The most sage advice that founders routinely forget is what concrete idea are you going to implement and why do you think it will work? There has be a way to invalidate your idea and as a corollary you must have the focus to collect the data and properly invalidate it.
Hey Mehul, crossed paths with you at AWS. Good to see you are doing your own thing now. We could connect sometime
I have had reservation about Ng from a lot of his past hype, but I thought this talk was extremely practical and tactical. I recommend watching it before passing judgement.
https://toolong.link/v?w=RNJCfif1dPY&l=en
not a single word about overwhelming replacement of humans with AI. nothing about countless jobs lost. nothing about ever increasing competition and rat-race. (speaking of software, but applies to all industries). his rose-glasses view is somewhere in between optimism-in-denial to straight-up lunacy. if this is the leader(s) we have been following, this should be a wake up call.
1 product manager to 0.5 engineers for a project? That seems... off.
strong MLM energy vibe in that talk.
You become a millionaire by selling books (courses) of how to become millionaire to others.
Relevant video https://youtu.be/CWMAOzH20mY?si=Kr8vp1vo_PpRNJ8-&utm_source=...
Thanks
[flagged]
A good chunk of Ng's work these days seems to be around AI Fund [0] which he explicitly mentioned in the video, in the first 5 seconds, involves co-founding these startups and being in the weeds with the initial development.
Additionally, he does engage pretty closely with the teams behind the content of his deeplearning.ai lectures and does make sure he has a deep understanding of the products these companies are highlighting.
He certainly is a businessman, but that doesn't exlcudethe possibility that he remains highly knowledgeable about this space.
He's lost credibility in my eyes given that his courses essentially have a pay to play model for startups like langchain
Except they aren't pay to play unless you consider doing the work for the course the "payment". There's certainly an exchange since there is a lot of work involved, but DLAI provides a team to help design, structure and polish the course and then the team creating the course does the majority of the work creating the content, but there's no financial exchange.
The DLAI team is also pretty good about ensuring the content covers a topic not a product in general.
The content is a repackage of previously existing, publicly available notebooks, docs, YouTube videos. I wouldnt be surprised if the repackaging was done by AI.
Courses are not academic journals, dude. They're supposed to be teaching you existing knowledge.
Again this is not true. I’ve known several people who have made courses for DLAI and they all put substantial time into creating the courses.
Baidu.
The video's description is about building startups through vibe coding, not using "AI" like self-driving or chatbots in startups.
Additionally, Baidu wasn't a startup when he joined in 2014.
Ng built baidu's AI department and began their start in various sectors with actual AI system design, so yes, he isn't a failed startup entrepreneur like any vibe startup maker who already wants to stop and give advice.
Maybe you can help me hire a vibe coder with 10 years experience?
He built it without LLMs in 2014 and now he is selling LLMs for coding to the young. That is the entire point of this subthread.
Right.. He's just a giant, not a midget with a step ladder.
But I do question why anyone who played a significant role in the foundation of the current AI generation would teach an obvious new Zuckerberg generation who will apparently think they are the start of everything if they get a style working in the prompt.
If not for 3 people in 2012, I find it highly unlikely a venture like OpenAI could have occurred and without Ng in particular I wouldn't be surprised if the field would have been missing a few technical pieces as well as the hire-able engineers.
He doesn’t have to at this point, he just throws money at younger ones that will build it.
I want an Andrew Ng Agent.
Not affiliated, but someone's already working on that for you: https://www.realavatar.ai/
I'm serious, the man's a genius...
... in essence, an "A-Ngent".-
(I'll see myself out ...)
He literally builds companies and hires ceos to run them Google it
> He literally builds companies
Like with actual mortar, brick by brick?
[dead]
when there is a gold rush, just sell courses how to mine gold
He sold courses (great ones!) long before there was AI-gold rush. He's one of the OG players in online education and I think he deserves praise, not blame for that.
[flagged]
[flagged]
[flagged]
I haven’t watched the video yet, but title does sound like quantity over quality.
Why faster and not better with AI?
I think this is an interesting question, and I’d like to genuinely attempt an answer.
I essentially think this is because people prefer to optimize what they can measure.
It is hard to measure the quality of work. People have subjective opinions, the size of opportunities can be different, etc, making quality hard to pin down. It is much easier to measure the time required for each iteration on a concept. Additionally, I think it is generally believed that a project with more iterations tends to have higher quality than a project with less, even putting aside the concern about measuring quality itself. Therefore, we put aside the discussion of quality (which we’d really like to improve), and instead make the claim that we can actually measure (time to do something), with the strong implication that this _also_ will tend to increase quality.
I think speed isn’t our problem.
Most of the time the problem it‘s quality but everyone only seems eager to ship as fast as possible.
Move fast and break things already happened and now we are adding more speed.
„Your scientists were so preoccupied with whether they could, they didn't stop to think if they should."
Or for the more sophisticated
https://en.wikipedia.org/wiki/The_Physicists
Energy consumption and data protection were a thing and then came AI and all of a sudden it doesn’t matter anymore.
Between all the good things people create with AI I see a lot more useless or even harmful things. Scams and fake news get better and harder to distinguish to a point where reality doesn’t matter anymore.
I think quality takes time and refinement which is not something that LLMs have solved very well today. They are very okay at it, except for very specific targeted refinements (Grammerly, SQL editors).
However, they are excellent at building from 0->1, and the video is suggesting that this is perfect for startups. In the context of startups, faster is better.
Depends on the startup. For medical or financial things faster isn’t better.
DOGE acts like a startup and we all fear the damage.
I would prefer better startups over faster at anytime.
Now I fear AI will just make the haystack bigger and the needles harder to find.
Same with artists, writers, musicians. They drown in the flood of the AI created masses.