[w10-1]

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))

[lubujackson]

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.

[davorak]

> 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.

[hakanderyal]

From the recent threads, it feels like the other half is totally, willfully ignorant. Hence the responses.

[rhizome31]

As someone who is part of that other half, I agree.

[fullstackchris]

> 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

[jart]

I like Andrew Ng. He's like the Mister Rogers of AI. I always listen when he has something to say.

[mnky9800n]

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.

[koakuma-chan]

Is he affiliated with nghttp?

[dmoy]

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.

[janderson215]

Wu from Wu-Tang?

[yorwba]

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

[57473m3n7Fur7h3]

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

[dmoy]

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

[stego-tech]

> 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.

[void-star]

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.

[vlovich123]

> 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.

[stego-tech]

> 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.

[vlovich123]

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.

[contrast]

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.

[handfuloflight]

This is great thinking, thank you for writing this.

[handfuloflight]

We've defined agents. Let's now define orchestration.

[ramraj07]

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.

[handfuloflight]

Okay. We've tried to define agents. Now let's try to define orchestration.

[lhuser123]

And make it more complicated than K8s

[jliptzin]

Not possible

[vajrabum]

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.

[lloeki]

> 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.

[nivertech]

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...

[lloeki]

> AOP always felt like a hack. I used it with C++

Colour me surprised that a language which is the practical definition of bolted-on features has an implementation that feels like a bolted-on hack ;)

> early on

Also the "early" AOP was very much focused on a formal definition and a specific approach, further entrenched by how Java did it.

But you don't need to do AOP "by the letter" to do AOP: as long as you have cross-cutting concerns that you sprinkle over your code you're doing AOP, even if you're not calling that pointcut and stuff.

Python decorators, C# attributes, Ruby's whole "monkey patching" (e.g alias method chaining pattern, or simply modules combined with the ancestor system and message passing) are all pretty much variations that fit well with the AOP paradigm.

I'd argue that a lot of people are doing AOP without ever realising it.

[alex_smart]

Also very much alive and called that in the Java/Spring ecosystem

[tomrod]

No need to add AI to the name, especially if it works. PaaS and IaaS are sufficient.