lukeschlather

I use profiles for this, I wonder what the tradeoffs are between the two approaches. I've actually been thinking about building some tools to make it easier to create and manage profiles on the fly. For me I really think profiles are better since the boundaries are harder - there's no worry about accidentally hitting the "log in with Google" button and clearly identifying myself to a site or similar, because I read the news in a separate profile from my Google account. And it would be nice if those sort of guardrails were very intuitive and easy to set up by default. But I feel like containers may be too squishy.

Did they ask any substantive questions? It sounds like they asked movie trivia questions, in a way that the LLM was almost guaranteed to be wrong. So extremely low-stakes where there is no conceivable downside to being wrong.

This sounds like something Claude Code could do very easily. If you need to actually look at the videos that's harder but still possible, but if it's just aligning GPS times and timestamps with reasonably accurate clocks, Claude Could probably do the ffmpeg commands unsupervised. I wouldn't be surprised if Haiku (the cheapest model) could do it, or an open-source agent harness with another small 30B model.

Just prompting claude (probably I would start with Opus) "I would like a HUD display of the following metrics from my Fit file overlayed on these GoPro videos, and I'd like the videos stitched together (there are some gaps, I want seamless playback) it would probably do it in 30 minutes or less, and the majority of the time would probably be ffmpeg.

I think Lockheed Martin is less leveraged than either SpaceX or Tesla. Also SpaceX just folded in several of Musk's companies that seemed questionably solvent. I really have zero clue how to value the resulting mess of liabilities and highly questionable revenue streams mixed with very solid ones, but $200B doesn't sound unreasonable. $1.6T sounds very high. My intention behind comparing to Lockheed Martin was to presume revenue would double for two consecutive years, which I viewed as reasonably charitable.

The article mentions uses of AI but doesn't really give any examples of harm from AI. It does give specific examples where it sounds like Kaiser is optimizing calls to minimize cost rather than improve quality of care. It's also pretty easy to expect that the examples given (treating longer calls as a problem and penalizing nurses for giving more than 3 pieces of advice) reduce the quality of care.

SpaceX's annual revenue is estimated at $40B for 2026. A comparison might be Lockheed Martin, with annual revenue of $75B in 2025, and a market cap of $120B. So I wouldn't say $200B sounds low at all, even if their revenue doubles (twice, since $40B is actually a projection where they were doing $20B last year, though it's apples and oranges with the acquisition of Musk's failing companies.)

It seems very plausible Signal deliberately curses itself somehow, given its extreme privacy focus.

That doesn't help you if you're the only person with the apk on their phone.

He didn't say he bets everything against ASI, he said he bets everything against ASI being a flash of light in the sky which destroys our chance of getting access to the wealth it creates.

Local ISP: “We’ll be happy to run fiber and new ethernet through your existing network conduits, trench to the curb, and help bodge in active poe ethernet repeaters for runs that are too long.”

I live in a major metro with a half a dozen apartments constructed within a block of me while I've lived here and this is very much not the case. I call them, they say they'll be happy too and then they ghost me. Of course I also can't get Starlink.

Using (inter)national labels when discussing local politics is incoherent.

They're looking at national politics, and on that score, yes, the ST is liberal. When it comes to local politics the Seattle Times is conservative. Your objection is a bit like people who object that the Democrats are conservatives, from a European perspective, except this is about the city itself. The current mayor is liberal, the previous mayor was conservative. They're both Democrats and would be defined as liberals by your factcheck, but that's a different rubric than the one we're talking about.

At the core this is a hardware problem. 1M tokens is simply not enough context to understand a codebase the way a human would understand it. Being able to selectively forget is potentially a very valuable power, but right now it's a substitute for a human's ability to remember the rough shape of something, decide it's uninteresting, and remember that it is uninteresting.

They talk about memory only being useful when guided by a human, I think the proper solution is deeper than that, it probably involves feeding the entire codebase and every agent session into a finetuning of the model, though at that point you might want some guidance to avoid feeding certain sessions into the model. Or maybe not, maybe the bitter lesson applies.

The Apple Vision Pro. I really wish Valve would release a $3500 Steam Frame with all the bells and whistles. Of course, with the price of RAM it might be $7000 so maybe I'll just have to wait until hopefully the market cools down a bit and baseline hardware advances some more.

Fable happened to be released after I had been experimenting with Claude Code for roughly two weeks. I had been trying to use Sonnet, and when I switched to Opus it was night and day. My understanding of geometry was maybe not as good as it should've been, and I kept seeing Sonnet say things I knew were wrong but didn't know enough about 6DOF camera positioning to ask it to fix. I finally asked the right questions, it couldn't answer them at all, I switched to Opus, it was night and day. But! Opus still couldn't really keep 6DOF "in its head." When I left it to its own devices it tended to come back having forgotten that it needed to keep 6 degrees of freedom in its head and collapsed the problem down to 3DOF or just a single angle.

Fable just understood what I was talking about and never needed me to stop it and say "you forgot this thing we talked about." The difference in spatial reasoning capability between the three models is very very palpable. I am curious to get more time with it because ultimately I feel like I sandbagged it by giving it problems that would've been within Opus' abilities, but required a lot more handholding.

what is the known-good thing? The whole point is that LLMs were not optimized at all, they got better results than older ML algorithms just because they are able to use all of the GPU, where older algorithms are designed for 10yo GPUs and can't make use of modern GPUs. But now you do in fact have to optimize, to the point that transformers look a lot more complicated than "attention is all you need."

Really? Here are 3 examples, I'm not sure what you mean, Congress has extremely broad taxing authority. If what I suggest is unconstitutional (I'm curious to hear your justification for that) they could probably do it by putting terms of use on the PACER data, have a paid tier specifically for orgs that resell the data.

* Sports gambling excise tax: https://www.irs.gov/businesses/small-businesses-self-employe...

* https://en.wikipedia.org/wiki/Federal_telephone_excise_tax

* Indoor tanning services: https://www.irs.gov/businesses/small-businesses-self-employe...

You asserted that we need the money PACER fees bring in - taxing businesses like LexisNexis seems like a way to get the same kind of money that wouldn't be regressive, if we really need to replace that revenue, and it comes from more or less the same place, without being paid by people who can't afford it.

We know where all the money in this system is: it's in LexisNexis and Westlaw. Each of them has revenue over $3B, individually. Presumably they have some lesser-known competitors. PACER fees are $150M. What percentage of PACER fees are paid by LexisNexis and Westlaw anyway in the course of their data ingestion? I'm not sure it really matters, we could simply restructure with a new tax on value-add services like LexisNexis that pulls in $200M, make PACER free, and everyone would be happier, probably including the value-add legal service providers.

The hardware to match Opus costs at least $200k, and you have to maintain it, and it's still not going to be as good.

What's a "basic machine learning system?" is this a question of the size of the model or the algorithm? Which algorithms are basic? If you've got an ensemble of models that includes multiple transformers not all of which are LLMs as well as CNNs, how do you think the marketing people should express that?

The R&D expenditure seems reasonable, and the revenue numbers seem realistic. I have no trouble believing they can be profitable by 2030 or much sooner. What I don't get is how you get from $30B in revenue to a nearly $1T valuation, but that seems almost level-headed compared to SpaceX, and it's not like any of the big tech companies' valuations make much sense in the context of their revenue.

The OS shouldn't be making many big changes that force me to reorient. When I'm moving between different UIs I often want to compare them; animations make it harder to compare state A to state B. I can detect very fine differences between two images by switching between them within a second, if there's a 1-second animation it not only means it's going to add a second, it adds a bunch of visual noise which might make it impossible to be able to distinguish what's an actual difference and what's just noise introduced by the animation.

Approximately nobody is throwing away phones because the OEM stopped providing security patches.

This becomes a practical reason more quickly than you think. If a company only provides 4 years of security updates and they only provide 2 android MV releases, you quickly become out of date. I had a BlackBerry Key2 that I bought in 2018, I had to replace it in 2024 and I was really holding onto it despite a lot of practical problems - Slack dropped support for the version of Android a year earlier, it was only when I tried to install Google Wallet and could not that I finally decided despite the hardware and software functioning fine it really wasn't practical to use a device that was stuck on such an old version of Android. (I would've tried to figure out the kernel myself if the bootloader wasn't locked.)

I have mixed feelings about Kindle, but I mostly read books on my phone these days, and my Kindle library is always there. I also have a physical bookshelf, but if I'm not home I can't review it so in some ways it's often less tangible than my Kindle library which I always carry with me.

I wonder how those who have experienced the pleasure of having 1000 children experience that compared to the first one.

Binaries are source code outputs, they are copyrightable and patentable. Weights are not copyrightable so people can freely extract the weights and run them. If Google patents any of the novel algorithms here releasing it all freely isn't an impediment to making people license it.

The flicker/signin redirects are a regression introduced post-Github acquisition. Really just evidence of the product being abandoned, or possibly of general operational decline at Microsoft which is affecting all their products. Definitely since the acquisition, both products have declined but Github has declined more to the point where I might prefer Azure DevOps.

Coding is a pretty small slice of the markets in play. Google's models are driving cars right now. Using coding agents doesn't give much insight into performance in the broader world; I would assume assume Google is performing better in general even if Claude or Codex is currently outperforming for coding.

I wasn't using Claude Code, but I told an agent to add something like this to the AGENTS.md, it did it and then a few minutes later it attempted to grant itself permission to do something and managed to delete the VM it was running on in the process. I have since adjusted the way I sandbox agents to make that less likely, but the moral of the story is clear.