mchusma

Okay, it's kind of somewhere between haiku and sonnet level pricing, at somewhere between sonnet and opus level performance. Its a great option. I was hoping to see opus class intelligence at haiku level pricing out of google, and this is close to that!

I think the original DeepSeek moment seemed important. And yes, the more recent model is good, but there are multiple. This commodification trend spans many different companies, including Kimi 2.5/2.6 and GLM5.1, and even Google itself with its Gemma models. There are a dozen models that exist at roughly the frontier from 6 months ago at 1/10th the cost.

Retatrutide is incredible in a mind altering way. I know 4 people on it, including myself. And all just love the way it makes you feel. You just don’t want to eat as much and you want to eat healthier. As far as I can tell, that is it. While I’m nervous there could be long term side effects, I think the health benefits of being at a healthy weight mean that for probably 50% of the developed world they would benefit from it.

Looking forward to more formal studies though!

I get this desire and commend the author, but I just want self driving cars and so I think we are just stuck with this.

I think this may be in fact a response to Starlink. Dead zones create opportunity for satellite direct to consumer plays, bypassing the big three. By teaming up around these fringe cases, you reduce that a bit.

It's funny that DocuSign was included here — small businesses really have no business using DocuSign. It's an enterprise product, it's expensive, and there are 20+ cheaper options out there.

eSignatures are probably in the top 10 at risk businesses right now. I did eSignatures before, and even I didn't buy a project just vibe coded the signature piece for a recent product.

I can't speak to their open source, but I really enjoy Bambu 3d printer and it works great for me and my 9 year old son.

What I will say is that this is probably the first model after gemini live to do some of these things. It feels similar to gemini live, which I don't think is what they were going for exactly, but IMO it is still impressive as I don't think anyone else has matched full duplex video/audio/tool calling.

Next gemini releases coming next week though, we will see how that matches up!

This paragraph is really odd: “As oppressive as it was, communism contributed by breaking down old social barriers and opening higher education to factory and farmworkers who had no chance before. A post-Communist boom in higher education means half of young people now have degrees.”

It feels like despite overwhelming evidence presented in the own article that communism was bad, they felt they had to say something vaguely nice about communism. But they can’t even keep it going for more than a sentence, because the next sentence says actually education was better after communism.

My father used to say "Nonprofit doesn't mean that nobody can make a profit". Seems applicable here.

I’ve looked at these auth providers many times over the years and I just don’t get the value. It takes me a couple of minutes to set up auth. Why would I want a dependency? It doesn’t help me with the hardest part which is configuring Google and Apple sign in stuff on Google and Apple. I just don’t get it.

I actually like the combined SpaceX / xAI as a hardware focused compute provider/infrastructure play.

I do think that if LLMs are the equivalent of more employees, then you'd expect people to do more managerial work and less individual contribution. Certainly, that's what I see in my day-to-day.

If you were to believe a lot of metrics Gemma 31B it’s much better than flash lite. It seems like I should be able to pay Google to use it and that should be at least a secretary, called action how I can do that but it’s missing from both the blog post entirely.

Part of the issue is Google complex web of products. There’s vertex Gemini Google AI studio Google edge. But I literally had trouble finding how to use this in my existing paid Gemini API account.

I find it puzzling Google doesn’t actively promote its own cloud for inference of Gemma 4. Open source is great, love it. But shouldn’t Google want me to be able to use and pay for it through Gemini and vertex?

Yes let’s go make deafness optional for adults and eradicate childhood deafness. Zero moral ambiguity.

Google owned 14ish percent of Anthropic before this investment, so presumably this could bring it up to as much as 25%?

I work with a lot of full-time devs, and it is very hard to go beyond the $200 max plan. If you use API credits, and I think the enterprise plan kind of forces you to do this, you can definitely incur this much, particularly if you're not using prompt caching and things like that.

But I and others in my company have very heavy usage. We only rarely, with parallel agentic processes, run out of the $200 a month plan.

And what do I mean by "hard"? I mean, it requires a lot of active thinking to think about how you can actively max it out. I'm sure there's some use cases where maybe it is not hard to do this, but in general, I find most devs can't even max out the $100 a month plan, because they haven't quite figured out how to leverage it to that degree yet.

(Again, if someone is using the API instead of subscription, I wouldn't be surprised to see $2,000 bills.)

At first, I was more excited about the Flash model, but I'm now more excited about the Pro model in many ways. I feel like the Pro model with an Run through unsloth, and with some fine tuning, is gonna be enough for many vertical SaaS applications.

Where previously I was wary to under-provide the intelligence level, I'm now more excited about the idea of being able to give these pretty large intelligent models to my application. The idea that for basically sub-agents, we can fine-tune them, should reasonably expect to perform as well as Opus for a specific subtask of which my applications have many.

In other words, we can run a general-purpose intelligent model, Sonnet or Opus, orchestrating a fleet of, let's say, 30 to 50 of these sub-agents that have been fine-tuned. By doing that, I can get very low pricing versus something that would have occurred if I used Opus or Sonnet for everything.

I honestly think health data should be public by default to any health researcher. We should do whatever we can to solve disease and live forever. Privacy be damned, I want life.

For comparison on openrouter DeepSeek v4 Flash is slightly cheaper than Gemma 4 31b, more expensive than Gemma 4 26b, but it does support prompt caching, which means for some applications it will be the cheapest. Excited to see how it compares with Gemma 4.

I appreciate this, makes me trust it more than benchmarks.

Kind of? But I really care about price speed and quality. If it used 10x tokens at 1/10th the tokens and same latency I would be neutral on it.

Kimmi 2.6 for example seems to throw more tokens to improve performance (for better or worse)

I’m pretty convinced that this is what Kimmi 2.6 does, mostly just thinks more.

I like both GLM and Kimmi 2.6 but honestly for me they didn’t have quite the cost advantage that I would like partly because they use more tokens so they end up being maybe sonnet level intelligence at haiku level cost. Good but not quite as extreme as some people would make them out to be and for my use cases running the much cheaper, Gemma 4 four things where I don’t need Max intelligence and running sonnet or opus for things where I need the intelligence and I can’t really make the trade-off is been generally good and it just doesn’t seem worth it to cost cut a little bit. Plus when you combine prompt, cashing and sub agents using Gemma 4, the cost to run sonnet or even opus, are not that extreme.

For coding $200 month plan is such a good value from anthropic it’s not even worth considering anything else except for up time issues

But competition is great. I hope to see Anthropic put out a competitor in the 1/3 to 1/5 of haiku pricing range and bump haiku’s performance should be closer to sonnet level and close the gap here.

You can run models on FPGAs and get massive cost, speed, and throughput gains (like 10x). The reason people don’t do it is because of other improvements (algorithmic) means that nobody really thinks locking into a model makes sense…yet. Would I want to use gpt 4o for anything today at 1/10th the price? That would be $0.40 per input, $1.50 per output. Gemma-4 31b is much more capable and cheaper. So a FPGA version of the model is just not worth it today.

But if progress begins to slow down, then the economics work. Maybe Gemma 4 is a good example. It feels really generally useful. Getting it at 1/10th the cost feels like it could be competitive in 2 years.

Closer to free than now…so I support. I would prefer us automakers to start competing, but at this point I’ve given up on Them.

I think one issue I have with Zone 2 proponents is that it makes sense that if you can do a lot more stuff at Zone 2 than you can at less. For most endurance athletes, the volume of running beats everything else in terms of efficacy. However, most of us don't have unlimited time.

For example, if I do a week of 3x30 minute swssions at Zone 2, my fitness is going to plummet. But if I do one at Zone 2, one at tempo, and one set of intervals, I'm at least maintaining fitness.

Would 5 hours at zone 2 be better? Absolutely. But I dont always have the time.

Ever since I saw the first one of these one-bit models made by Microsoft, I thought this was a fascinating route. I assume that in practice, this is less helpful than it seems, just because there's every economic incentive in the world for the big AI labs to produce small, powerful, fast models. None of them seem to be using this technique, so it's interesting, but I suspect it's not quite working.

I also have yet to see any of these at a larger scale. For example, can you try one of these at 100 billion parameters?