lukeschlather

That is a setting. The first time you get refused it offers to stop or switch to 4.8 and if you choose 4.8 it automatically downgrades silently whenever it happens again. I just shut it down and switch to Codex when this happens.

I would also agree that 5.5 feels dumber the past week. More weird logical problems. it's not as dumb as 5.1 or 4.8 but it feels like it regressed.

The thing is max reasoning is most effective at things which are least interesting (the ones where the answers can be validated by a computer.) I've found for a lot of tasks like image recognition, low reasoning can even be better quality than max.

What do you think? Required for what? I don't think we should be landing on the moon unless we can build a sustainable presence there. I think the only way the other moon lander is better if it turns out to be a better path to getting more equipment on the moon over time, and my feeling is that's unlikely. I would actually be somewhat surprised if Blue Origin landed their small lander on the moon before SpaceX lands Starship HLS.

I could see China being "first" but their first isn't anything resembling a sustainable presence.

I would call Blue Origin setting off the largest launchpad detonation in history a lot of things, but I would not call it unserious.

I'm excited at the prospect of Starship landing on the moon. 100T payloads are maybe not required to make it happen, but probably are. For those of us who care about Artemis IV happening on schedule, being disappointed in a 52t payload capacity is very straightforward.

I'm not excited about more Starlink satellites.

> Apple’s iPhone will not autocomplete references to drugs, sexual acts and, perhaps in the spirit of corporate protection, “iMessage isn’t working”.

It will be interesting as corporate predictive text generators take over the world and companies like Apple will literally build the Jobsian reality distortion field into their products to portray their services as working better than they are.

Yes, Carbon and Mythos are the secret super-smart models you can't use without an invitation and a million dollars.

Carbon is to Argon as Mythos is to Fable.

Yeah, why didn't I think of just spending $30k/month on an inference server and figuring out how to train a 2T parameter model. I clearly don't care about this at all or I would just do that.

I think we've reached a point where reading the AI summary (which I think is Gemini Flash Lite) is probably going to be more reliable than actually reading the top 10 search results. The top 10 search results were often blogspam even in a pre-AI world. I feel like either you trust Google to retrieve information or you don't (and Gemini has problems but so does Google Search, and are you really sure the latter is worse than the former?)

It has been over a decade since Google Search was anything but a machine-learning powered word association engine. I miss old Google Search but it's dead and I trust Gemini more than what Search has turned into.

AWS and GCP both work pretty well if you avoid conditions, which are a nightmare. Azure is like everything is conditions, and also buggy.

I'm sure there are absolutely brilliant people working at Azure. I think the overall system is very badly architected and I don't know what it would take to salvage it. Infra security is a lot easier to do when you have a system that has mostly solved it already, and I'm not saying GCP and AWS are perfect, but "security is mostly solved" is a decent description of their systems, I think. Of course it could all fall over tomorrow, but I'd guess that's a lot more true of Azure.

If Intel were capable of building more chips they would. I don't think their lagging TSMC is a diabolical attempt to inflate their profit margins. Although I only really know what I read on the news.

The Danish status quo is radically anti-capitalist and anti-business by American standards. Do you think the average Danish zillennial is in favor of moving to a more American model (privatized healthcare, minimal union protections, etc.) ?

I had Claude write me a script to probe all the GCP Vertex models. I can say with confidence that I know how to call Argon, because Claude told me how, I did it, saw it was modestly better than Flash for my use case and I decided it wasn't worth the money right now, for this use case.

I'm fairly confident I can't call Carbon, and I am okay with that.

The problem with that is then you would have GPT 6.1 Max with max thinking which is different from GPT 6.1 Small with max thinking.

If you look back the naming problem started around the point that OpenAI started doing reasoning models (o1, 4o) where they had to describe this fork in model behavior that let a smaller model think longer and be smarter.

You can come up with naming schemes that don't clash, but there's something to be said for this approach where the thinking level is like you describe and the model name requires you to learn that Astra is the largest model.

There's obviously a limit to what you can change via ToS. Yes, their ToS says they can alter the deal at their discretion. Adding ads to a premium ad-free tier seems like fundamentally a violation of the presumed contract. (Specifically talking about one-year plans, saying that they can make these kinds of fundamental alterations to the deal mid-year sounds like it ought to be illegal if they're selling one-year plans.)

That's fair, changing mid-year is definitely a problem and arguably illegal under current law, though really if we had functional antitrust I'm not sure that bundling would be legal to begin with.

That seems like revisionist history. Amazon started out with Prime Video bundled with Prime as a whole, and didn't raise the price. After 5-6 years they made it an add-on, then they added an ad-supported tier.

But if anything, the racket is bundling free shipping with ad-free video and everything else. Raising prices periodically to keep up with inflation is not a racket. And whether you would prefer to get downgraded to the advertising plan or automatically pay more is a personal choice, and automatically downgrading you rather than renewing you at a higher price is in some ways a more sensible default. (Though I'd prefer to be upgraded to the higher price generally and be notified so I have the option to downgrade or cancel, though downgrade really isn't an option I would consider.)

The margins on GPUs are estimated at 80% right now. So the price would most likely drop by 70% just by building the additional manufacturing capacity. Assume some modest reductions in cost and the price should drop by 80-90%.

The only reason computers are so expensive is that demand is so out of control.

This seems like an AI-generated summary and it leaves some obvious questions unanswered.

This thread (linked by the page) seems like a better source: https://groups.google.com/a/mozilla.org/g/dev-security-polic...

Ian Carroll notes:

> I think it's worth noting that this compromise is much worse than the private keys simply being stolen by a researcher. They are inherently weak single-byte keys that anyone can factor, and the issue cannot be contained.

I would bet that there are some proofs whose simplest form are simply incomprehensible to humans. Explaining them in terms a human could possibly understand is a fool's errand. However the theorem itself may be much simpler, and the theorem is the thing that's most worth translating.

Attempting to translate the proof is valuable, but we have to be realistic about what is and isn't possible.

It's not a constant factor, that's the misunderstanding. It's an unknown factor.

I've had pretty good luck having Gemini or Claude use SAM as a tool. Gemini is really good at identifying things in images, but it doesn't do a great job doing segmentation per se, but SAM on its own can't really decide what you need to segment.

They're looking for like, plumbers' email and quickbooks which tells you what kind of plumbing jobs they do and how much they get paid, so they know what kind of things to make robots for.

Most people don't use password managers. I am not even sure that password managers really qualify as a reasonable amount of complexity for the average person. There's a cost/benefit and I think passwords + SMS 2-step-auth is a reasonable balance between usability and security for most people.

Obviously, I use a password manager and I use MFA where possible, but even that I find to be a little questionable from an availability standpoint because I have a hard time ensuring that I have multiple MFA devices located in different buildings, so a fire or a flood or something could totally lock me out of all my accounts, some of them possibly permanently, which is the thing that scares me more than getting hacked.

Passkeys (combined with full-disk-encryption by default) scare me because they make the assumption that most people would prefer to lose access to their photo library than risk someone breaking in and getting them, and I actually think most people would prefer the converse. If a burglar breaks into my house they can steal my TV, I can live without it, but please don't steal my photo albums (problem is they're the same thing.)

It's not an exploit, it's a gap in our understanding of the algorithms. We know it's optimal, but we don't know the Big-O complexity.

I'm not sure I see the difference between a complex pairing dance and needing to install the iCloud Passwords app on Windows (I've used iTunes for Windows.) Writing down a password is much more reliable than expecting any of that to work.

Bluetooth is very hard to use. This just isn't about hardware support, it's about "how long does it actually take the average person to do a Bluetooth-mediated workflow, how consistent is it between devices, if someone is trying it with a new device, what is the learning curve?" If someone can't figure out how it's supposed to work between their iPhone and their Windows laptop, there might theoretically be support. Practically speaking it's broken most of the time.

The deeper problem is that the default behavior is storing a passkey directly on a specific device and hardware-locking it to that device. That's great if you're an enterprise or whatever and you've got SSO set up well. For the average person, "I lose my iPhone and it effectively resets all my passwords" is a hopelessly unusable workflow.

That's fine if you've got a couple of supported devices. The hardware support is however very bad. If my hardware is unsupported (and most hardware in the world is unsupported) that's not very helpful.