hijodelsol

While I understand both of your perspectives, please think about how frustrating it can be to have poured years into building a product genuinely solving a problem and building a relationship with clients only for a dozens AI copy cats to come in. Even if your product is clearly better, the SEO slop alone will raise your marketing cost, price cutting will affect you as not all clients are able to tell quality before purchase or will know of your project, etc. And in the end, everyone is worse of. We get less for the products we build. Customers in many cases get a worse solution. And the AI copy cats multiply until none of them can make a living from it.

I would think that for training Google has optimized their entire setup for TPUs so I can’t see how this would be used for anything but inference

There are definitely more shades of grey. On my iPhone I can select a close contact to be able to overturn my protection but this contact needs to have security features turned on, too. So Apple staff cannot do it, only a non publicly known person that has 2FA and encryption themselves. Add time delays, notifications, identity checks and more to it and you can make this process reasonably secure while still ensuring recovery.

As explained in another comment, I think this is more about Google orienting Flash towards more complex use cases. If we got minor improvements vs 3 Flash with 1.5x the price so they can optimize their margin (which on such small models for conversational tasks is a completely different stories than the 3-25x subsidies that these agentic coding plans offer) I would have been happy. Or even no change at all. But knowing Google, I now must fear that they will deprecate 3 Flash without offering any realistic option for that user facing chatbot segment that does not require multi-tool use across 500k context.

When using Hetzner, DigitalOcean or any other VPS service together with Cloudflare, I can handle millions of page views for 5-50$ a month at pricing that has stayed nominally the same for a long time and due to inflation and performance gains of the underlying chips has basically become cheaper.

I am aware that it was likely subsidized or at least did not have appropriate margins. But over time, that same capability should become profitable if parameter efficiency and chips improve. For many customer facing use cases outside of coding assistants, optimizing for speed, basic logic/maths and conversational texts matters much more than being able to use 40 tools simultaneously. I would have hoped that Google would recognize this and keep a dual line up, where Pro and Flash are clearly intended for different market segments. But it seems, it's all in on coding assistants and screw the other use cases..

Now, we might need to change to DeepSeek 4 Flash if Google deprecates 3 Flash.

It's discouraging to see Google price Gemini 3.5 Flash at 3x the cost of Gemini 3 Flash. I would think that most people that deployed this model in production would have used it for low latency tasks, classification/categorization, customer support or basic RAG-/RAG-style chatbots. Performance on coding benchmarks is nice and all, but where is the "intelligence too cheap to measure"? This new cost point is quite prohibitive and will eat up a lot of margins if developers adopt it.

I mean, the 100$ plan is less than the hourly rate of any consultant / senior dev in developed countries. So if it can save even one hour a month, it's cost efficient for the customer (at the current, subsidized rates, of course).

I came here to say something similar. As someone who works in a field that applies machine learning but is not purely focused on it, I interact with people who think that arXiv is the only relevant platform and that they don't need to submit their work to any journal, as well as people who still think that preprints don't count at all and that data isn't published until it's printed in an academic journal. It can feel like a clash of worlds.

I think both sides could learn from the other. In the case of ML, I understand the desire to move fast and that average time to publication of 250-300 days in some of the top-tier journals can feel like an unnecessary burden. But having been on both sides of peer review, there is value to the system and it has made for better work.

Not doing any of it follows the same spirit as not benchmarking your approach against more than maybe one alternative and that already as an after-thought. Or benchmaxxing but not exploring the actual real-world consequences, time and cost trade offs, etc.

Now, is academic publishing perfect? Of course not, very very far from it. It desperately needs to be reformed to keep it economically accessible, time efficient for both authors, editors and peer reviewers and to prevent the "hot topic of the day" from dominating journals and making sure that peer review aligns with the needs of the community and actually improves the quality of the work, rather than having "malicious peer review" to get some citations or pet peeves in.

Given the power that the ML field holds and the interesting experiments with open review, I would wish for the field to engage more with the scientific system at large and perhaps try to drive reforms and improve it, rather than completely abandoning it and treating a PDF hosting service as a journal (ofc, preprints would still be desirable and are important, but they can not carry the entire field alone).

There are ways to independently fund open source projects, though. I have previously contributed to the Python Software Foundation and to individual open source maintainers through GitHub donations (which are not dependent on GitHub, as there are many alternatives). Projects like the Linux Foundation exist, too. And government funding, especially for scientific endeavors or where software is used to fulfill critical state tasks, is an option, too. I refuse to subject to the hypercommercialization of software and still believe in the principles behind open source.

This might be true for uv and ruff, and hopefully that will happen. But pyx is a platform with associated hosting and if successful would lock people into the Astral ecosystem, even if the code itself was open source.

They could have joined projects like the Linux Foundation which try to not depend on any single donor, even though complete independence from big tech is not possible. I don't know the motivation behind Astral's approach, but this acquisition does leave a weird taste behind about how serious they were about truly open source software. Time will tell, I guess. (Edit: typo)

This is a serious risk for the open source ecosystem and particularly the scientific ecosystem that over the last years has adopted many of these technologies. Having their future depend on a cap-ex heavy company that is currently (based on reporting) spending approx. 2.5 dollars to make a dollar of revenue and must have hypergrowth in the next years or perish is less than ideal. This should discourage anybody doing serious work to adopt more of the upcoming Astral technologies like ty and pyx. Hopefully, ruff and uv are large enough to be forked should (when) the time comes.

This did happen in summer of 2022, but only in public buildings, private households were not affected, so the OP's point seems a bit overly dramatic. Given that AC usage is highest when solar production is also highest, this seems highly unlikely given the solar build-out of the last 4 years.