janalsncm

Yes, it reads as self aware without moral conviction.

Lex, I don’t know how to put this nicely. But if you are reading this, sell your teslas. Finger wagging about Twitter/X has roughly zero moral weight coming from someone who has bought two Teslas.

The bumper stickers are worse, unfortunately. They read as self awareness without moral conviction.

> Huh? What "moral panics" proved what exactly wrong?

I think they are referring to moral panics around violent video games and the idea that shooting a player in a video game they teaches people to be violent in real life.

It seems like you are making an analogous argument, that being mean to Claude teaches people to be mean in real life.

In my opinion, being an asshole has a way of self-correcting. People will not want to hang out with you anymore if you suddenly browbeat them at a dinner party. So I am not that concerned about this.

> Permitting abuse and cruelty might carry real misalignment risks.

Please expand on this. What are the specific risks of swearing at Claude?

I don’t understand why Anthropic would want to convince people that LLMs are conscious. When people think of “country of geniuses in a datacenter” does he want people to imagine a datacenter filled with slaves?

Yeah I’m not trying to argue it is AGI, but it’s not as simple as a short script. There’s some amount of debugging involved, and no amount of if-statements could cover all possible ways a script could break.

In a way, “recursive self improvement” just means tools helping us to create better tools. At least that’s what the words mean.

A bit too pessimistic imo. I agree that AI can’t automate things end to end, but a good deal of R&D involves kicking off a training run and babysitting it.

If your training run dies at 1 am and you’re sleeping, you won’t find out about it until the next day. You can lose up to 18 hours of work depending on when it happens. Based on the error it might be as simple as tweaking a single hyperparameter and rebooting, which is something LLMs are usually capable of.

Even just that task means I can kick off multiple runs over the weekend and have confidence they’ll finish. It’s a game changer.

In my experience R&D has basically two axes: how innovative it is, and how well we can measure the results.

For the quadrant of non-innovative tasks where we already have a good way to measure performance, Claude can handle this. There is very little ambiguity, and we are basically just looking to maximize some metric under a set of constraints.

Many business processes are not like that. They might be conceptually simple, but it isn’t that easy to say whether a system has done a good job or not. I would say that LLMs can help with this a lot but they have bad judgement because it requires talking to people.

And the other, perhaps more rare issue is in problems where there is data but actually modeling it to sufficient quality or fast enough is hard.

When buying our refrigerator, I didn’t look into how evil the CEO of whirlpool is. It’s quite possible he is as bad or worse. However, at least he has the good sense to be quiet about it.

> isn’t the United States a free market

No, far from it. A free market doesn’t have regulations, subsidies, tax incentives, bailouts, etc.

Ok, let’s say it leads to 60% offshoring and 40% hiring more Americans.

That 40% is a win for Americans. It means more new grads will be hired instead of losing out to “experts” from TATA to copy and paste from ChatGPT. Our vibe coders will be home grown, corn fed.

The argument works no matter what you think the breakdown is, 70/30, 80/20 etc.

Hiring remote workers also doesn’t require filling out H1B applications and dealing with USCIS. If using remote workers was just as good, companies would already do that.

Agreed. I’m specifically responding to the statement that Haiku was higher on all benchmarks. Normalized by cost (and why wouldn’t you normalize by cost?), it was higher on 1/3.

Excited for their IPO just to put this to bed, one way or another.

If we look at their performance per dollar charts,

In OSWorld 2.1 Haiku is better.

On GDPval-AA v2.1 Haiku is equal or worse than Luna.

On Humanity’s Last Exam they don’t seem even be comparing Haiku with Luna.

For these baby distillations of flagships, I expect their users to be very price sensitive.

The first thing the user sees should directly answer the question they asked. Not a different question. They didn’t ask what a roast looks like.

The user is probably in a grocery store. They need to buy the stuff first. Showing them a picture of a finished product is an irrelevant distraction.

Then, if there is some arguably relevant content you can tack it on afterwards.

I was about to comment the same thing. Why is a picture of a roast helpful in that moment? The user asked for a plan to prepare a meal. They need a list of ingredients not an image.

I would say these all boil down to empathy. Think about who your target audience is, and write for the least informed among them.

It’s ok to be selective about your target audience. Most of us are writing for free anyways so we’re not losing revenue by not explaining what a computer is in an article about optimizing LLM throughput. You might be writing to other engineers who are familiar with the topic but not the particulars of your project.

Put the most important thing above the fold. If you catch someone’s attention in the first 10 seconds, it buys you another 30.

Add visuals. Boxes and arrows, charts, videos where appropriate.

If you must write a meandering narrative, put it at the end, not the beginning.

> Maybe someone can tell the LLM's about these anti-patterns, like, seriously, would it help?

In a way, by posting this article online the author has already done that.

If that’s true, he doesn’t need free legal advice from us.

A lot of lawyers will give a free consultation, and in my experience (not for IP) they will give decent expert advice for free. No harm calling them.

Because describing something in words will never have the fidelity of directly applying the changes yourself.

I was referring to the Deepseek R1 paper, but there might be more recent research. I hadn’t heard anything about Fable reasoning stability.

I think the more intuitive mechanical explanation is, in RL when you are assigning rewards to a rollout you might give a reward for stable reasoning and another for correctness.

If you are just summing the two, a rollout with better correctness can score equivalently to a rollout with a better answer. So ultimately you can end up with worse answers.

I’m already paying the real unsubsidized price with my own GPU.

Maybe it can’t solve a Millennium Prize, but it’s pretty good at writing CSS.

> Accessibility has always been a foundation of web development

> CSS should be hard

I don’t see how the author can square these two. If accessibility is important, the CSS which enables it should be easy. And what about accessibility from a web dev perspective? If doing the right thing requires arcane knowledge do you think most people will buy a book on CSS to master the thing that AI is actively de-skilling?

In my opinion, the author needs to take their own advice. Talk to actual people who need web dev work, figure out what they want to pay for that can’t easily be done by machines, and get good at that.

Before RL we typically SFT on human reasoning traces. This makes the reasoning traces somewhat coherent and the model trains faster.

But you don’t have to do that. You can skip straight to RL. If you do, the model will generate complete garbage reasoning traces before generating the correct answer. In fact, if you add a coherence reward to the reasoning trace, the model will perform worse (since you’re now diluting the correctness reward).

Imagine we did that, split up a model layers as A->B->C. C will need to wait for B to compute a forward pass, which is waiting for A to compute its forward pass. To compute the forward pass, B needs all of the outputs from A, which is an upload and a download (maybe these can be done concurrently).

Then A waits for B to compute its backwards pass, which is waiting for C to do the same thing. Again you are sending around potentially gigabytes of data.

This is in contrast to mining bitcoins for example which doesn’t require any coordination from miners because their work is completely independent, and the answer is very small compared to the work needed to get it.

> colos are very hands off

That’s a pretty broad statement? Maybe a more correct one would be that some are hands off enough that some people might find them worth it.

And that really only addresses one problem. The other one is the coordination problem. You go in on a $30k GPU with 3 other people. So each person puts down 7.5k. But then one person wants to pull out, and so every remaining person needs to put in an extra $2.5k on top, and they’re questioning whether $10k for a chatbot is really worth it.

Well this is why I think small local models are the opportunity of a generation. Because it’s a bit transgressive these days to say you don’t need a big centralized service that spies on you and requires a subscription. It’s not the absolute best, but it’s yours.

I agree in general that we technical solutions to social/political problems are suboptimal. But technical solutions don’t require the political system to work, which is a very attractive property at a time when politics seems broken.

There is a counterpoint that retreating into technical solutions only cedes more ground and makes things worse, and for that I don’t have a good reply other than the fact that life is short.

Knowing nothing about this, I wonder if it could be useful in situations where we can’t reliably sync with all the workers. Something like folding@home, where all the workers are just shaking weights and if one of them finds a winner it uploads to the central server?