ashdksnndck

Congress has mandated that USPS has to be self-funding. If the leadership at USPS fails to raise revenue, they are forced to cut services. You could write a letter to your senator.

If the low version of the model keeps putting out bad output, it can go through more iterations at the harness level before getting to a solution.

The world does not have a finite supply of “deserves sympathy” juice which must be carefully metered. One can reasonably feel sympathy for people (and animals too!) who are experiencing different forms of suffering, even if some are much more severe than others.

If you’re trying to distribute aid resources or something, it would be worthwhile to get an understanding of who is suffering the most. But if someone is just having the normal human experience of feeling sorry when they hear another person is hurt, there is no benefit to intentionally cultivating schadenfreude to shut that down. That’s not an instinct we should encourage in each other.

I assumed most domestic sales were going to closed frontier models, through some mechanism or another (direct ownership, cloud resale, whole-gpu or whole-dc lease, etc). Got any numbers?

Ultimately if you have a GPU that can fit X parameter model at Y raw tokens per second, you want the smartest weights. If the frontier closed model is Z% smarter or “more valuable” you would be willing to pay up to Z% margin for the right to use it. Hence whether you are running open or closed model depends on pricing strategy of the frontier labs, also whether the frontier labs are able to diffuse their own models better than open labs can.

If I have a really straightforward question and I want a fast answer I’ll sometimes Google search for the AI overview. But it’s so unreliable. I use opus 5.5 for any question that Google could plausibly screw up. This is a huge product failure on Google’s part, IMO. I’m a $20/paid user for both Gemini and Claude. It’s nuts that with Google’s massive head-start in AI (and apparently they still are capable of making a frontier model), they somehow deliver such a crappy version of AI by default in their flagship product. Google is probably in some part responsible for the public’s distaste for AI.

Easy way to tell will be to wait and see if it gets fixed now that there is attention on the topic.

> if Anthropic models run in a Google cage, does that not also mean Google can distill Anthropic models even easier (both fast and unseen)?

Anthropic is indeed trusting Google and other companies they rent compute from. Distilling or stealing the weights would be a contract violation, and would wreck other business relationships. I would be less worried about Google doing this, and more concerned if there is some sketchy neocloud in a van offering suspiciously low pricing.

They’re giving 0% interest, which is a discount. It’s rational to take that instead of paying cash.

It makes sense to use an LLM to write the queries when exploring data, but if you are building a dashboard that’s going to be used repeatedly, you should save the queries so they can be validated and improved over time. The LLM shouldn’t be rewriting a new query every time you open the dashboard.

Interesting mindset.

You want to drive across town. You, only willing to deal in certainties, must decide ahead of time whether you will get in a crash. If you determine won’t get in a crash, you make the drive, pedal to the metal. If you determine you would crash, you stay home.

Me, willing to deal in probabilities, make the drive at the speed I feel reasonable and wear a seatbelt.

I don’t believe there is any fundamental reason LLMs can’t do this task, and the base models are probably already capable of it. RL to elicit the right behavior just hasn’t been done yet. Coding got solved first because it was cheaper and easier to get the necessary data for RL environments, which made it the natural first market opportunity. Solving white collar communication will require more resources to collect the data and build the environments, but the labs now have more resources and are working on it.

I want to believe reasons why this won’t work, but it feels like cope. It would be smart to be prepared for the world where knowledge workers get replaced by AI and also prepared for the world where AI is a bubble that pops.

The goalposts on what neural nets can possibly do have already moved past chess, go, protein folding, coding, poker, math, the Turing test…

Deceitful vs delusional or incompetent matters for your legal defense, your parole hearing etc

Was anyone confused by the title and thought we are exporting more gold in volume or weight than other products? It seems obvious that gold must be winning on value.

They still had to engineer the demo unit.

I demand a refund!

So if you knew in advance about the timing of both the tech bubble and the real estate bubble, you could have made a lot of money trading on that?

The input is unlimited size and the algorithm is achieving a speedup with constant size code.

> What industry are you in?

Technology. Don’t what to be too specific about where I work.

> Was the team disbanded specifically because of your productivity with AI?

Company has laid many people off, and moved many into new lines of business. I don’t know how management decided who went where. I don’t get the impression I’m a special case or anything.

RAM manufacturers are bidding against NVIDIA and everyone else for the same constrained supply of EUV machines. And it takes years to build more fabs. Micron has multiple fabs coming online in 2027 and 2028.

My employer is paying thousands per month for my tokens. Multiply that by a million people and there’s your $50B per year.

You might ask: how could this possibly pencil out for my employer? Answer: I’m the only person left on what used to be a team.

The margin call will be brutal if you’re wrong and it takes a few more years than you thought for the trend to break.

Hmm, I wonder if the difference is that I’m working on an existing large codebase that wasn’t originally written with AI. It might be picking up practices from the surrounding code. It’s happily using the same crates that are already there, and occasionally bringing in new ones. Some of the code is “weird” but I also learn from it.

The Rust I’m getting from Opus 5.5 works great. What’s the complaint?

Actually I felt since 4.5, it was “good enough” for basic coding (and competing models have reached that level as well). The big improvements since then have been longer-horizon tasks. Giving it a vague description of a problem, then come back hours later and find it did a better job than I would have. And since 5.5 I haven’t encountered any more load bearing seams.

Maybe it’s slow, but I’m not interacting with the models in a live chat anymore for most work. I’m giving them tasks to do async. The limiting factor is actually my bandwidth reviewing their output.

I wonder if people downvoting consider your question to be a thoughtcrime or something. It’s a legitimate line of inquiry. Research happens because somebody is interested and writes a grant request that is accepted. Governments and institutions where the research happens have rules as well. It’s possible that the researchers, the folks reviewing the grants, and the people making the rules haven’t done a cost-benefit analysis that adequately covers the tail risks. For some kinds of research, the answer can be: no, we shouldn’t do this. It’s reasonable to publicly discuss these concerns.

The top city of origin for Burning Man attendees is Reno.

It sounds like there was a personality conflict or something in the specific camp you went with? Some of the issues you describe are totally foreign to my experience and don’t seem descriptions of the event as a whole.

The only lawyer language I can remember was the warning on the ticket that I might die. I don’t think I ever encountered any contracts or similar.

Couldn’t the labs time-shift training and other batch workloads to make up for regular changes in inference demand?

In 2024 (pre-trade-war), the US exported $168M of wine to the EU and $460M to Canada. Any recovery in wine exports to the EU won’t make up for Canadian boycotts.

You don’t need to get to 100% every night if your vehicle has hundreds of miles of range. If you’re charging 40 miles overnight and using 20 miles on average, and your vehicle has 250 miles of range, that gives you massive margin to do stuff that consumes lots of additional miles on random days. In the rare edge case that you actually use up your entire range, you can then go to an L3 charging station and top off quickly.

Obviously, if you’re the kind of person who often drives 100+ miles day after day, then L1 charging isn’t going to meet your needs. Those people should budget for an L2 charger at home. Otherwise, upgrading your home charger is a fun optimization for nerds, but not a prerequisite to own an electric car.

Why do you care how long home charging takes? It’s not like you have to hang out in or around your car while it’s charging, the way you do when filling it at a gas station. As long as your car charges enough overnight to make up for your average daily usage, you’ll be good.