mchusma

i think as the pricing has gone up on the Chinese models it has made them less appealing, and with the introduction of Gemma-4 not many are at the pareto frontier (also in my experience, not just the stats): https://arena.ai/leaderboard/text/overall?viewBy=plot

it feels like between K2.6 and GLM5.1 we have Sonnet level intelligence at roughly Haiku level pricing. Which is great.

I'm hoping that Anthropic will be able to release an updated Haiku soon and they really need something that is 1/3-1/5 the price of Haiku to compete with the truly cheaper models (Gemma-4 is really good at this range).

You should do what you want, and as a break it’s fine. But IMO right now the most leverage for most people is learning how to effectively manage agents. It’s really hard. Not many are truly good with it. It will be relevant for a long time.

Anthropic has been avoiding the hard thing, but they just need to do SOME kind of pricing thing here to shift demand. I expect there is just no amount of tricks that can handle the few hours at peak load. They need "surge pricing".

This seems reasonable surge pricing approach to me: 1. Implement surge pricing for everyone for the peak 2 hours of the day if possible. 2 hours you can work around, 5 hours is too hard. 2. Give existing customers a one time credit for surge 3. Make sure the plans just consume credits at an accelerated rate (e.g. if on the max plan, i just get 1/2 usage during peak hours). 4. Exempt sonnet/haiku from surge (so people can keep using) 5. Make "auto" settings in claude code etc automatically adapt during surge hours, so people don't get surprises by default. 6. For the first 90 days, unofficially waive the fist $100 in surge for every user but notify them. To train users about the surge, and get them used to it without having them actually pay. 7. (I don't think they would do this but this would help) allow users to fall back to using something like GLM 5.1 or Gemma 4 automatically in outages, with a partnership to handle it. Its not ideal, but i would prefer it in "partner mode" than not. IMO they can charge like 10% on top of the partner fees for this if used in other times, but during outages or surge, partner mode is free. But 100% managed by Anthropic so users don't need to set things up and can just use the Anthropic harness.

Maybe they should not allow routines during peak for now? might help shift load.

I do think the Iran crisis should continue to push countries towards nuclear + solar. Like Ukraine helped shift some in Europe back to supporting nuclear after foolishly shutting down reactors.

developer jobs are up: postings ~11% YoY, tech workforce +1.9% projected for 2026, BLS 15% long-term growth. Rebound from 2024-25 lows

Jevons paradox

I agree this is helpful, and the numbers are better. I think this is one of those situations where everyone is a bit right: 1. Its a meaningful jump 2. Its not quite as extreme as some people will say. Opus might have gotten there with more tokens, so capability might already be close to in range (maybe) 3. A pause to let people review security postures is conservative and probably good (is this 30 days? 90 days? 180?) 4. Anthropic is so compute constrained that they probably couldn't handle a full rollout right now anyway 5. If you thought claude code with opus was basically AGI, you probably think Mythos is AGI. If you thought it was far off, Mythos is probably still off for you.

I did a bit for fun and felt there were areas where I did do well. However, the problems with these markets is the contracts are often such where you can win or lose on obscure technicalities. I bet on 5 things, and 4 were resolved in ways I felt were non obvious.

For example, will the Ayatollah be out by April. Lost because death apparently did not count. Bet Anthropic would advertise at the superbowl. Lost because while they did so they advertised Claude not for the company.

Once it became clear that it was an exercise in legalese, I was done.

Without knowing much this seems kind of reasonable? Some baseline of care but not enough that most people will opt it.

If I could make universal healthcare in the US it would be something like: - free generic meds where there are at least 3 independent low cost providers. - free ai diagnostic tools - tax breaks on healthcare spending - catastrophic insurance @ fixed prices for various diseases (eg if you get colon cancer you get $30,000 for treatments. Basically whatever the low end of reasonable is, to invent you to price shop etc.). These numbers should keep being adjusted down over time as people get more efficient at treating. - maybe up to $400 per year on various preventative tests.

And remove most other regulations on healthcare (including banning employer insurance).

If the government provided this, the cost for private healthcare would plummet. People would have free access to diagnostic and drugs better than anything you had 20 years ago (2006). Which is pretty incredible!

I think teslas issue is that they need the AI5 chip for robotaxi ops, the current chip just doesn’t cut it. So if they have batches end of 2026 and start optimizing the models, by mid 2027 volume production you might have robotaxis coming online at about 100k per month. Waymo currently has less than 10k cars on the road.

Lots of ifs here. If they can enable hardware 4 for robotaxi ops then they can have 3m+ cars ready to go. But I am skeptical of it. And given that Elon’s top priority is scaling chips and AI5, I think that is proof that he thinks it is likely necessary too.

So 1m robotaxis by end of 2026 is theoretically possible but I think unlikely, and it’s more likely in the 200k-1m by end of 2027. If they pull that off, they could still be largest by then if Waymo doesn’t rapidly scale. Fun times!

I think the way I and others use it is code with clause, review or bug hunt with codex. Then I pass the review back to Claude for implementation. Works well. Better than codex implementation and finds gaps versus using Claude to review itself in my opinion.

When people talk about peptides, they typically mean either (1) GLP compounds like Reta that have successful phase 3 trial data and will be approved or (2) stuff like 157 which has no evidence or plausible mechanism to work.

I’m expecting the new iPhone release this fall to be coupled with some great version of Siri/model. This could be the first reason I’ve seen to upgrade in a while (although even that I’m not sure of, as I am king of in the “always use the best model it’s worth it” camp.)

Apple has a great shot at making a highly optimized 4.5 version of this model highly tuned to the next gen iPhone, which could work great.

Feels like it is a “supply chain risk” using the same logic Anthropic was labeled such. (Note I do not think any should be a supply chain risk, but if they were being consistent)

Doing a bit more research, this looks like it might perform roughly as well on text tasks with modest context windows, so may be just a better cheaper option unless you need a million token window.

For those curious, on openrouter this is $0.14 input and $0.40 output, or ballpark half of Gemini flash lite 3.1 (googles current cheapest current gen closed model)

I believe the biggest benefit of going to space, particularly in building space stations, is making humanity focused on building a bigger pie.

This is one step towards this. But once we can build (effectively) infinite land, we will be in true abundance.

I was excited about the 1.58 bit models from a year or two ago, but the never seemed to go anywhere. Curious in particular how this scales up.

I hope it’s much longer. Bunch of 200 year olds having needed to work to 75. Crazy talk!

curious if zjit will be ready to be default over yjit next ruby release. hope so!

Not sure I buy the premise. Austin has a household $133k median income with a $435k median home price. It’s very affordable.

But, to make Austin more affordable still, you make it less expensive to build so that it’s profitable to build. Typical regulations that do this are: - Lower minimum sizing requirements - open zoning - raise height limits - make sure you don’t have unwarranted restricted fire codes (some places have elevator stairwell requirements that are insane) - make permitting easier or not required at all for some cases - no min parking requirements

Pretty sure as good as Austin is, they could easily reduce the costs by up to 30% (there are parts of the country with 50% the cost per sq ft for new construction).

It’s disturbing to see multiple comments that “capitalism” is to blame for the costs in healthcare. The US healthcare market is the least free market in the US. Capitalism (when there is true competition) brings down costs pretty universally. Those parts of the market that are market oriented tend to get better and cheaper over time (for example, generic over the counter meds, needles, and even health care panels, AI health advice) all getting better and cheaper. You can get pretty cheap panels of blood drawn, for $200 or less with tons of data. It’s only when you hit the regulatory wall and you exit the market do you see costs explode.

Our imperfect system pays for the worlds medical R&D, so I would actually love to see per capita spending remain similar BUT have the market opened up, with a nice safety net at reasonable cost, and money pouring into curing aging and all disease.

I occasionally use wireless, but 90% of the time I prefer wired for this reason. Reliable, simple, harder to lose, no charging, easier to connect. I just done think the wires are a problem for me, sometimes a benefit, sometimes a mild inconvenience.

Having tried this a bit I do really like the single api call for all of it.

I also appreciate transparent pricing but I am not 100% sure the sense of scale of costs. It could be helpful to give some ballparks on things for each of the plans. I'm not sure exactly what i could get out of a plan. My guess, trying hard to figure it out, was if i had about 1,000 pages of new/updated content per month, I would pay $295/month for unlimited queries on top of it. Is that roughly correct?

I think to the extent they are making a speed v quality tradeoff, I think they are making the right call. 10x speed over quality any day for me. Reminds me of:

"If brute force doesn't work, you aren't using enough of it." - Isaac Arthur

I like that it lets you specify the types of accepted docs. The biggest issue i have with Stripe identity verification product right now. And biometric re-log in is also great. Will check it out.

I’m skeptical.

The average cost of car ownership is $0.69 per mile without insurance, $0.25 per mile to store it, and $0.49 per mile in societal costs (death, injuries, delays due to accidents). So about $1.43 per mile. I do not enjoy driving, so would add more cost per mile, maybe some would want to pay more but I do t see that much joyriding outside of teenagers and classic car enthusiasts, so I don’t think those that do it for pleasure is a large population.

Tesla cybercab is targeting $0.20 per mile. Waymo projections are $0.40 per mile by 2030. Assuming both hit $0.50 and are twice as safe, this is basically $0.75 per mile.

I don’t see may paying more to drive themselves. And I think as society there will be non economic reasons human driven cars get banned. Like MADD but for human cars.

So I expect 5 years and human cars will not make sense in many cases, 10 years new human car sales to be <50% current levels, 15 years you start seeing bans. 20 years bans common.

Red light running is bad...but I think the solution to this problem at this point is just "self driving cars". With some exceptions, I would just focus all jurisdictions on this future and avoid policy inline with a world full of self driving cars. Currently in the US, most places feel like you need a car, and many US laws are designed with this in mind. In 5 years, this will no longer be true, so laws should reflect:

1. No parking minimums 2. Less free parking (e.g. street parking) 3. Policy supportive of self driving cars 4. More aggressive removal of driver licenses for human drivers with repeat violations 5. More aggressive penalties for driving without a license.

Conductor (mac app) does some of this, might want to take a look.