mchusma

I think the signal from Anthropic is pretty clear between Haiku not getting an update in a year and the Sonnet issues this year. They don't care about low intelligence models. You should go elsewhere.

That's what we've done, migrated workflows away from Haiku and Sonnet. I actually think this is not a crazy position because these lower models have so much competition from Grok, OpenAI, DeepSeek, and about 20 other labs with really solid models in the Haiku to Sonnet range. So what is the point of Anthropic competing in these spaces where everything is going towards zero cost?

I’m just not convinced that most artists would change their behavior under no copyright, so somewhat agree. But the main issue is length of time.

The fact that copyright is longer than patents makes no sense. Patents prove that a shorter period is sufficient to stimulate large investment. And you have to pay for patents!

If you assume reasonable discount rates, the economic value plummets over time for an asset producing the same income (a bad assumption for copyright work in general).

At an 8% discount rate, extending copyright from 32 years to infinity adds less than 10% to its total present-day value. There are many good reform proposals, but mine would be something like 15ish years of free copyright, followed by 1 paid renewal for 15 years that is modest in price (eg $1000) and one extra 10 year renewal that is something like $1M. Meaning you have most economically neutral work free up at year 15, small business types having it affordable through 30 years, and the big Hollywood productions can go from 30-40 years for a modest (to them) fee to incentivize investments in expensive works. Movies come to mind. Give existing work a 20 year buffer on top of that to transition and there is a very small economic effect on everyone.

The government could always pay for results too, like $10B for something that does this. I don’t love it either and am not positive it is better than patents.

I would agree that there are a lot of people in this space, both products and roll your own. We are doing our own stuff here in our company right now, and could be interested. But for me, I didn't feel like i know enough to actually assess and take the next steps. Good luck!

This is all solvable with abundance. We could put together nuclear power with solar and desalination and carbon removal.

Let’s start acting more like a kardishev II civilization.

> If it's so much cheaper to build multiple small reactors, just build one big plant with 24 small reactors.

The biggest costs to nuclear are associated with each of them being unique snowflakes. They need to be standardized and mass produced to bring down costs.

So the dream is many big plants (eg starting 10+ per year), which is what France did and China does, but since we can’t seem to have that here, small reactors are an attempt to solve that.

You realize his examples are public sector unions. Which mean the “employer” you are referring to is the citizens? Not some evil corporation, government mandated monopolies which then get taken over to serve their union members.

I don’t have any objection to voluntary unions. Meaning people can join them for support or collective bargaining, but that citizens can just opt to not employ anyone in the union. Unions could be support structures for workers, instead they are mafia bosses.

I have an agentic workflow and Luna just always gets stuck, SOL and grok 4.6 don’t. I like Luna in theory I just find not much practical work for it yet in coding type work.

Now I think Luna is plenty good for many applications inside a very good harness/scaffold. And I think there are a lot of those usecases. So I think these small models are really good for application developers.

But for entrepreneurial knowledge work all of my work still benefits a lot from more intelligence.

FDA is just too slow. Whole I’d like to do more radical reform, at a minimum this stupid long wait should be tossed. If someone hits a target endpoint auto approval pending further review or something

Remember in COVID when results. Came in and they took a month to schedule the meeting?

Some of this has to do with limits from the ACA (Affordable Care Act), which limited the margins of insurance companies. It creates incentives for higher premiums, but also these types of gains, which is just bad for everybody.

I don't think there's much you can look at with the Affordable Care Act and think that it was a success.

Wild numbers. I already thought GLPs were huge but this makes me think it’s even larger.

You are correct. I think this is the bull case. It seems like this would be useful right now for some things (eg moderation).

Maybe! (1) Would SOL level intelligence be useful 3 years from now? 5 years? (2) would dedicated chips be the most affordable way to run this model in 3-5 years?

I suspect the answer to both of these questions is yes right now, but I agree it’s borderline.

Yeah, the Data Center NIMBY crowd is making these look real good right now.

I think they talked about this being general purpose chip but I would think that Anthropic/OpenAI are at the scale now they could bake LLM weights into chips themselves.

For example, GPT Sol baked into a custom chip run for $100M that runs 10x as fast and 10x as cheap should pay for itself as long as the chip is useful for long enough.

While 2 years ago nothing was useful more than 1 year long, there are many older models in use now (e.g. Haiku 4.5, GPT-OSS 120b), and I expect this trend to continue.

I know this is what Taalas was doing (acquired by AMD), here was their demo, https://chatjimmy.ai/ which is based on Llama 3.1 8B. It feels like this should start to happen soon.

I’ve never met someone who didn’t think tipping was too widespread and too high. You say a majority of people are happy with the state of US tipping?

One note on patio11’s opinion on this is that he really overweights the ongoing work and innovation required for electronic payment processing. It WAS a great novelty and deserves to have made a lot of money for 30 years. But the reason they make so much money today is monopolistic low behaviors to lock in their advantages. It’s not a free marlet because of deals over time, some of the most famous of which are their prohibition on charging different rates for cards or even disclosing the rates on cards.

I think the most simple piece of legislation to solve a lot of problems is to allow merchants to pass along the interchange rate to their customers. If they could do this legally and operationally, this would solve most issues here. If a credit card wants to be expensive, fine the consumer should pay for it. Because of contractual and operational limitations, credit card companies have gotten themselves into the current arms race.

If stripe implemented this, it would make me appreciate them as a force for good instead of being a part of the problem.

I am excited for this! A great practical format.

I love progressive rendering. At 15% loaded in this example it’s surprisingly good already.

I was more skeptical of openrouter a year ago. Not hard to implement multiple models, but after getting burned multiple times (eg Google in particular) I see openrouter as serving a few key functions: (1) aggregating demand across neoclouds for performance etc (2) providing a great developer experience (3) easily switching models and normalizing the idiosyncratic nature of each. There are so many (good) model providers today this is more compelling now with 10+ solid model companies than a year ago when I thought it might be just 3.

So I have shifted load away from calling some models to openrouter.

I really think ebikes, and things like this, would have a much better reputation if they really only supported up to maybe 10 km/h or 6 mph. This is a normalish running/jogging pace. Basically, they helped up hills and when tired but were not "motorcycles-lite". At least around us, in an area with lots of kids, people want to have them but there are so many injuries and problems that sentiment is generally very negative. And I think all you need to "make biking easy" is low speed assist. You don't need support up to 20 mph for that.

I also had low energy but that was really mostly when I was losing weight, which made sense. As I slowed weight loss and reduced dose, energy improved.

Maybe? But GLM just put itself on the pareto today again, so its up to 10. The theory was that there would be recursive self improvement in models, leading to a 1 or a handful of entities running away from the rest. But basically the opposite happened? Different teams, different hardware stacks. What do we make of this? (1) There must not really be any deep secrets/moats right now? (2) improving models is not something you can throw only intelligence at now?

This is a solid release (at the intro pricing, the other pricing is dumb). I do think its a missed opportunity to really blow things out of the water and have this be another 1/2 off, but clearly they don't have the inference efficiency for it. Speed is good, the knowledge in google's models is solid for those usecases, price is reasonable (after 3.5/3.6 major missteps).

The intelligence index vs cost pareto frontier is crazy now, its basically a flat line with 9 people all at or right at the edge of the frontier along various parts of the graphs. Insanely competitive right now.

I think in this case its a sign they are subsidizing usage, so it means that the only people left using it in Jan 2027 will be some enterprisy or orphaned workflows which should pay up or migrate off.

Nice to see SpaceX on the model frontier! They have been chasing it for a while.

Many good comments here. It’s somewhat common for me to voice record say a blog post of product updates, more like a ramble. Then have Claude clean it up. Then argue back and forth about certain things until it’s good, then make a final pass sometimes to change a few key words. This is incredibly different than pure ai text. Presumably it will show as ai generated here, even though I would argue it is not really. So I can’t use Claude for this usecase anymore.

I think the solution is assume everything is ai generated unless told otherwise and rely on authorship/brand as a sign of quality.

Cool! I wish there was a way to print boards for hobby purposes. What is the lowest price you would work with? Ballpark

I genuinely think this is a good idea if implemented well and there is a human backup. I’ve had 911 not answer before. I’ve called for things that I wasn’t 100% sure are valid.

As long as implemented well.

Yes, that is the really compelling thing here IMO. Its a viable deepseek competitor for many people, and I missed that on the first pass.

This is a nice release and a solid improvement over Spark 1.1. It compares favorably with Grok 4.5. Not SOTA, but solid releases. I think they need to really get this more competitive with Deepseek V4 Flash / Luna pricing to move the needle.