zurfer

tldr: the author argues it is closer to costing 500 USD per month IF a user hits their weekly rate limits every week.

Which is probably a lot more correct than other claims. However it's also true that anybody who has to use the API might pay that much, creating a real cost per token moat for Anthropics Claude code vs other models as long as they are so far ahead in terms of productivity.

whenever i worry that AI will eventually do all the work I remind myself that the world is full of almost infinite problems and we'll continue to have a choice to be problem solvers over just consumers.

LLMs before extensive RL were harmless. Now with RL I do fear that labs just let them play games and the only objective in a game is to win short term.

Please guys and girls at those labs be wise. Don't give them counterstrike etc. even if it improves the score.

Yes, but honestly what's the best source when reporting about a person? Their personal website no?

I think it's a hard problem and I feel there are a lot of trade-offs here.

It's not as simple as saying chatgpt is stupid or the author shouldn't be surprised.

to be fair, i think usage has increased a lot because of coding agents and some things that worked well for now can't scale to the next 10x level.

6x more expensive

we use codex review. it's working really well for us. but i don't agree that it's straightforward. moving the number of bugs catched and signal to noise ratio a few percentage points is a compounding advantage.

it's a valuable problem to solve, amplified by the fact that ai coding produces much more code.

that being said, i think it's damn hard to compete with openai or anthropic directly on a core product offering in the long run. they know that it's an important problem and will invest accordingly.

I certainly don't think people on HN are dumb, I'm surprised that the sentiment towards this is just talking so much about the downside and not the upside.

And look I do agree that humans should be the one responsible for the things they prompt and automate.

What I understand is that you let this lose in a folder and so backups and audits are possible.

I'm a bit shocked to see so many negative comments here on HN. Yes, there are security risks and all but honestly this is the future. It's a great amplifier for hackers and people who want to get stuff done.

It took some training but I'm now starting almost all tasks with claude code: need to fill out some word document, organize my mail inbox, write code, migrate blog posts from one system to another, clean up my computer...

It's not perfect perfect, but I'm having fun and I know I'm getting a lot of things done that I would not have dared to try previously.

"Nobody" is against open source.

People are against bad products, high maintenance, high complexity... .. or against working for free without much reward.

It would be great if there was a way for humanity to have and develop free open source software but it's not obvious how to do it well.

It's an interesting industry that needs billions to bring a new drug to market. At the same time it creates a lot of value to a patient. But the manufacturing of a single dose is usually tiny.

Now how do you price that? The profits here will reward pharma investors and enable more investments in RnD of new medicines. I feel that's mostly fair.

Yes it's a bit disappointing but probably captures the current American and Chinese opinion quite well.

Europe as a whole has a lot of good things going for it but I do agree that it's less ambitious on average than these 2 power blocks.

However the same dynamic that was described in the article where nobody wants to lack behind is also true for Europe.

Also, yes Novo Nordisk plundered their GTM in the US and lost market valuation but you can still get the same medical outcome in Europe as a patient based on a European invention. Another one: The first Covid vaccine came out of Germany.

More interestingly is the question on degrowth. I personally believe that growth is the more tempting path in general, but we do live on a finite planet and no system is on a path or has a good framework on how to grow sustainably or responsibly. Maybe AI is going to figure it out for us, but maybe it involves some hard tradeoffs that intelligence alone can't solve.

You're too kind. Even the CEO of Google retweeted how well Gemini 2.5 did on Pokemon. There is a high chance that now it's explicitly part of the training regime. We kind of need a different kind of game to know how well it generalizes.

Yes I tried it with minimal and it's roughly 3 seconds for prompts that take flash 2.5 1 second.

On that note it would be nice to get these benchmark numbers based on the different reasoning settings.

It's a cool release, but if someone on the google team reads that: flash 2.5 is awesome in terms of latency and total response time without reasoning. In quick tests this model seems to be 2x slower. So for certain use cases like quick one-token classification flash 2.5 is still the better model. Please don't stop optimizing for that!

How do you understand meritocracy? It seems natural that those that do valuable things get rewarded a lot.

Ideally everyone would get the same chances to do valuable things but that's not how the world is setup. Unfortunately.

However trying to change that must be done with care as it's easy to increase injustice (looking at most communist systems)

It drives me a bit crazy when people say OpenAI has no moat.

Yes, companies like Google can catch up and overtake them, but a moat is merely making it hard and expensive.

99.999.. perc of companies can't dream of competing with OpenAI.

I think it's a great article that should discourage a lot of people to waste resources.

To really do it you have to treat this article as a to-do list of challenges to overcome. If you have no ideas on how to address those challenges you should not start.

well, maybe no one felt informed enough to write this, so it was outsourced to the llm (imposter syndrom) or it was pure laziness.

There are shareholders/owners and CEOs. You can certainly have an AI CEO if the board of directors wants that. Although depending on the jurisdiction CEOs might need be humans, but surely not everywhere.

And you could even imagine AI owners with something like Bitcoin wallets. So far it wouldn't work because of prompt injections but the future could be wild.

It also tops LMSYS leaderboard across all categories. However knowledge cutoff is Jan 2025. I do wonder how long they have been pre-training this thing :D.

I heard the story once on how you migrate these old systems: 1 you get a super extensive test suite of input - output pairs

2 you do a "line by line" reimplementation in Java (well banks like it).

3 you run the test suite and track your progress

4 when you get to 100 percent, you send the same traffic to both systems and shadow run the new implementation. Depending on how that goes you either give up, go back to work implementation or finally switch to the new system

This obviously is super expensive and slow to minimize any sort of risks for systems that usually handle billions or trillions of dollars.

You're right, I didn't produce evidence in my comment. Smoking and lung cancer are more clear cut than social media problems in general.

The effects of social media are more complex and nuanced than smoking. There are a lot of studies that show that social media has a negative effect on mental well being. When someones dies of loneliness or birth rates collapse and young people have less sex than ever, social media is never the only cause and might not even be the main reason, but it seems to play an important role.

Also, I don't think all social media is bad. I do love these discussions on HN even or especially when we don't agree, but tiktok and similar services have a lot of bad incentives with regards to user well being.

I don't buy it. The most relevant critique is see is that it's hard to control the age of your users without removing anonymous accounts thus limiting privacy. Well, it's a hard problem but it doesn't feel impossible to solve.

To me there is no question that children should grow up protected from harmful substances. You don't want kids to smoke, scrolling algo feeds is not better. There is enough interesting internet out there without social media!

The issue is rather the algorithmic feed optimized for grabbing our attention. It's definitely addictive and should be regulated like other drugs.

Give people technology, but let's have an honest conversation about it finally. As a adult it's already hard to muster enough self control to not keep scrolling.

The current go to solution for the kinds of problems that TabPFN is solving would be something like XGBoost. In general it's a good baseline, but the challenge is always that you need to spend a lot of time feature engineering and tweaking the data representation before something like XGBoost can deliver good performance on your regression or classification problems.

For me the promise of foundation models for tabular data is that there are enough generalizable patterns, so that you need less manual feature engineering and data cleaning.

And kudos to the team, I think it's a really creative application of neural networks. I was always frustrated with neural networks, since they were hard to tune on "structured" data and always under-performed (for me), but we also never had real foundational models for structured data.

this was the article I had in mind, when writing this: https://dynomight.substack.com/p/chess

Yeah I was not precise; it was `gpt-3.5-turbo-instruct`, other variants weren't trained on it apparently. https://dynomight.substack.com/p/chess

It makes me wonder if we'll see an explosion of purpose trained LLMs because we hit diminishing returns on invest with pre training or if it takes a couple of months to fold these advantages back into the frontier models.

Given the size of frontier models I would assume that they can incorporate many specializations and the most lasting thing here is the training environment.

But there is probably already some tradeoff, as GPT 3.5 was awesome at chess and current models don't seem trained extensively on chess anymore.

Well, I can't comment much on Genie, but the core question is always how you scale the complexity.

In Dot, it's divide and conquer. If you have several different teams each of them has to maintain their knowledge base.

A bunch of our customer have less than 10 tables hooked up to Dot, but this data is core to their business and so the analytics agent is really useful. Our most complex setup is on more than 5000 tables, but that was a lot more work to lay out the structure and guidelines.

Also, I don't think all organization are ready for AI. If the data model is a huge mess, data quality is poor and analytics use cases are not mature, it's better to focus on the fundamentals first.