428 karma · joined February 28, 2021
Socials: - bsky.app/profile/https://bsky.app/profile/curious-cat-163.bsky.social
---
Ultimately, individual habits do add up. But with climate, one would be hard pressed to find evidence that conservation is the path forward. It does not work, unfortunately.
"On one of these toy problems, the Tower of Hanoi, the models structurally cannot output enough tokens to solve the problem — the authors still took this as a claim that “these models cannot reason” or “they cannot generalize.” This is a small scientific error."
"it appears that a majority of critiques of AI reasoning are based in a fear of no longer being special rather than a fact-based analysis of behaviors."
Yes, the more recent generation of GPUs optimize for attention math. But they are still fairly "general-purpose" accelerators as well. So when I see papers like this (interesting idea, btw!), my mental model for costs suggests that the CapEx to buy up the GPUs and build out the data centers would get re-used for this and 100s of other ideas and experiments.
And then the hope is that the best ideas will occupy more of the available capacity...
But, for models, this is an interesting finding because a lot of LRMs are LLMs with a _bunch_ of post-training done on top. We know this about DeepSeek R1 (one of the models evaluated in the Apple paper) for sure. They write extensively about how they took DeepSeek-V3-Base and made R1 with it. [1]
If the post-training is resulting in lower performance on simpler tasks then it ought to inspire more research on how to make it so that it doesn't -- i.e., with more training (of any kind), we should be gaining more capabilities. This has been a problem with DNNs historically, btw. We had these issues when fine-tuning text/image classifiers as well. Some weight changes can be destructive. So, it has to be done with a _lot_ of care. And, I am sure folks are working on it, to be honest. Maybe some of them will say something here. :-)
Very clever, I must say. Kudos to folks who made this particular choice.
> we identify three performance regimes: (1) low complexity tasks where standard models surprisingly outperform LRMs, (2) medium-complexity tasks where additional thinking in LRMs demonstrates advantage, and (3) high-complexity tasks where both models experience complete collapse.
This is fascinating! We need more "mapping" of regimes like this!
What I would love to see (not sure if someone on here has seen anything to this effect) is how these complexity regimes might map to economic value of the task.
For that, the eval needs to go beyond puzzles but the complexity of the tasks still need to be controllable.
The question is do they still need 10? Or 2 would suffice? How about 5?
This does not need to be a debate about the absolutes.
I am curious -- what would have made it more effective at conveying information to you? Different people learn differently but I wonder how people get beyond the hurdles of jargon.
I'll offer my take as an outside observer. If someone has better insights, feel free to share as well.
In market terms, I think it is because Google, Microsoft and Apple are all still trying with varied success. It has to be them because that's where a big bulk of the users are. They are all also public companies with impatient investors wanting the stock to go up into the right. So, they are both cautious about what ship to billions of devices (brand protection) and cautious about "opening up" their OS beyond that they have already done (fear of disruption).
In technical terms, it is taking a while because if the tool is going to use LLMs, then they need to solve for 99.999% of the reliability problems (brand protection) that come with that tech. They need to solve for power consumption (either on edge or in the data centers) due to their sheer scale.
So, their choices are ship fast (which Google has been trying to do more) and iterate in public; or partner with other product companies by investing in them (which Microsoft has been doing with Open AI and Google is doing with Anthropic, etc.).
Apple is taking some middle path but they just fired the person who was heading up the initiative [1] so let's see how that goes.
My two cents.
[1] https://www.reuters.com/technology/artificial-intelligence/a...
Not sure exactly how you used Claude for this, but maybe try doing this in Cursor (which also uses Claude by default)?
I have had pretty good luck with it "reasoning" about the entire codebase of a small-ish webapp.
That may or may not be true for use-cases that require asynchronous, bulk inference _and_ require some task-specific post-training.
FWIW, my approach towards tasks like the above is to
1. start with using an off-the-shelf LM API until
2. one figures out (using evals that capture product intent) what the failure modes are (there always are some) and then
3. post-train against those (using the evals)
But, now they have network effects; > 30 million users -- that's a big market to build for.
Would love to hear folks' experience around "managing" all this new code.
However, shouldn't we ask for more? Even writing the paragraph above feels exhausting. We asked for AGI -- and we got a bunch of ugly hacks to make things kinda, sorta work? Where is the elegance in all that?
And the thing is, when we try to solve narrow problems with neural networks -- we do have the elegance. AlphaFold, AlphaGo, Text Embeddings, etc. All that stuff just works.
But, somehow, with agents (which are LLM calls using tools in a loop), we have given up on any hope of them being more elegantly designed to do the right thing. Why is that?
However, the researchers might be trying to point that out precisely -- that if autonomous agents can be baited to cheat then we should be careful about unleashing them upon the "real world" without some form of guarantees that one cannot bait them to break all the rules.
I don't think it is fearmongering -- if we are going to allow for a lot more "agency" to be made available to everyone on the planet, we should have some form of a protocol that ensures that we all get to opt-in.
Curious, why do you use it?
It is okay to feel like that for a bit. But, this story that you are telling yourself is not going to help you if you keep on telling this. Whenever I have felt negatively about myself like that in the past, it has helped to stack some wins -- even if little ones -- to get out of the rut.
So, if you want to try to be an engineer, start building things using tools/tech that you are interested in and put them in the open source. Or better yet, help out some open source projects that you respect and have benefited from in the past, if you can. Hiring managers (and I have been one) look for real evidence of skills mentioned in the resume and a well-rounded Github repo increases the likelihood of a callback in any kind of market. Stack some little wins there.
OR, if you want to be a product person, do the equivalent of stacking some wins in that realm. I have very little experience in that space so maybe someone else here can identify a concrete thing or two that you could do to attract hiring managers in that line of work.
> This example is in a format based on recommendations from Anthropic for use with Claude Projects. This works so well that we’ve actually found that Claude can provide even better information than our own documentation!
This is the file: https://docs.fastht.ml/llms-ctx.txt
Excellent analysis and I largely agree. The part where he looses me here is the certainty in the claim around fundamental changes to the economics...
This is open weights with a (largely) open recipe. Arguably, Meta has already been at it but now there are two labs producing frontier models at this level. A wishful analysis could imagine that there _might_ be more and that _could_ result in a qualitative shift in how the frontier model research is being done and hence affect the economics but Dario seems to have largely ignored this aspect.
Not sure why?
This is a problem that the Computer Science departments of the world have been solving. I think that the "good" departments already go for the "broad knowledge" of theory, systems with a balance between the trendy and timeless.
There is project-oriented engineering work that lasts for some arbitrary near-term period as well (2 week scrums, or 6 week long cycles in the shape-up world) -- but there is some deep strategic thought given to overarching design, etc. that I think the original post was talking about. Those don't fit neatly into OKRs and I buy that.
I would think that some marketing work is project-like but there have to be things that are more strategic in nature... Things that have impact beyond a quarterly campaign, etc.
That's where I was coming from. Feel free to disagree. Would love to get your take. :)
Just thought, I'd share what I think about the substance of the idea (not the implementation). I think a big untold story in the US healthcare system is how it shifts the burden of coordinating care to patients and/or their loved ones.
To be sure, there is a lot of decisions that the individual (or their NoK) should be making but the amount of paperwork that flies around and lack of coordination between say an insurance company and the provider is astounding. This becomes very pronounced for every corner case and the entire machinery is wired to record things in myriad systems but somehow not make things better when it comes to the core outcomes -- providing healthcare. Every entity in the food chain is out to (and does!) make a buck. Meanwhile, there is a wait time of > 30 days to meet one's primary care physician over a video chat!
So, I absolutely LOVE your idea. The implementation probably requires a lot of iterations here. One suspects that there are ways in which a consumer facing app could make some real money to level the playing field in favor of the patient while being a sustainable busienss.
I suspect they'll follow up with a full paper with more details (and artifacts) of their proposed approach.
I am from the engineering side of things but this does not seem right about marketing...
Also, the examples of OKRs for engineering in the article seem somewhat contrived. I am sure some organizations/teams have OKRs like "ship the roadmap" but to say that is what all of them look like (everywhere!) is generalizing too far.
Like, everything in life, there is probably a balance: Between thinking in the near-term vs. thinking about the long-term. Between thinking what the ICs on the front-line think is most important vs. what the "middle-management" thinks will move the numbers on metrics that the execs care about.
A well-run planning process will result in the right balance being struck. And that IMHO is a rare thing: a well-run planning process, i.e. My $0.02.
-> A new benchmark emerges with a novel evaluation method.
-> A new model saturates the benchmark by acquiring the novel “behavior.”
-> A new benchmark introduces yet another layer of novelty.
-> Models initially fail until a lab discovers how to acquire the new behavior.
Case in point: OpenAI tackled this last step by introducing a paradigm called deliberative alignment to tackle some of the ARC benchmarks. [1]
Alongside all this technical iteration, there’s a parallel cycle of product iteration, aiming to generate $ by selling intelligent software. The trillion $ questions are around finding the right iterations on both technical and product dimensions.
It was okay -- did a project that simulated some financial statements. I recall it being quirky but, to most humans (i.e. non-programmers), any programming language would be quite quirky so I don't understand the stigma around it.
I also learned C around the same time and that one stuck. :-)