IMHO a bigger worry is if they were opening the case then they could have modified it for spying or changing contents. I'd have thought stuff like this should have case alarm notification and inventory change alerts.
That doesn't make sense. If you want to take the food angle then the other serves homework. If you want to take the fraud angle then the other facilitates dietary fraud.
To be fair your dietician/healthcare professional might also be disappointed, but you're ultimately right, just like with college nobody else cares. Of course, that was already implied from the onset. I find it intriguing that you've only just come to that realization.
If you watch the Dwarkesh interview with Dario, he is extremely conservative with compute allocation as of early 2026, where Sam has been compute-pilled for at least 3 years[1] and believes that we’re going to have several OOMs more FLOPs in the near future than we have today.
You can also get a drift of it in their announcements - OpenAI has been far more aggressive in datacenter build-out partnerships as a more top-down allocation strategy since 2024, and Anthropic has rented out existing compute from infrastructure-heavy companies that are losing the model race in what seems like a more desperate ad-hoc compute acquisition strategy.
In all, Dario’s conservatism was in service of not going bankrupt, but any inspection of his argument that an overallocation of compute a year too early results in bankruptcy falls flat: there’s so much demand for mechanized intelligence in the world today that it’s easy to rent out compute if you have too great a supply.
[1]: recall that he was trying to court the saudis into investing $1T to build his own chips
Anthropic being very conservative about purchasing compute was a core theme of Dario's spot at NYT Dealbook in December[1]. At the time, he cast it as being wise and waiting to see where the chips fall. I'd say he might be regretting that position a bit now.
> According to the memo, which was reported by Bloomberg and CNBC, OpenAI put its own 2025 capacity at 1.9 gigawatts — a figure it said was three times what it had the year before — and placed Anthropic's equivalent figure at 1.4 gigawatts. OpenAI told investors it foresees its own footprint climbing into the "low-double-digit range" of gigawatts within a year and hitting 30 gigawatts by 2030, and that Anthropic would top out somewhere between seven and eight gigawatts before the close of 2027.
True, and the pricing plus token efficiency seem to suggest that gpt-5.6 is much more efficient. But hard to know since neither lab publishes that info and it’s a bit of an apples and oranges comparison.
They were famous for committing to buying compute with future money that many people thought they would go bankrupt. Anthropic was afraid of going bankrupt and didn't do the same.
There's less clear info about this from Anthropic's side, but given their frequent issues and lack of resets and overall less usage given to users on their subscriptions plans, it's a pretty simple conclusion to draw.
extremely well known fact in the industry. they have reams of spare compute for prosumers which is why they can afford to give away so many resets and have even more subsidized usage limits
we are in the "millennial lifestyle subsidy" era for AI where companies ruthlessly undercut each other in an attempt to win marketshare, before then ratcheting up prices
I find LLM's too fall into the trap of writing a bash script for a task that clearly needs to be implemented in an Actual Language with Real Data Structures. For example, ask an LLM to bring up a SQL server with some schema + data preload step, and it will write a profoundly long bash script to do that task, every time.
my experience as well. claude produces functional but over-engineered scripts fairly frequently with often very questionably useful safety checks, especially for powershell.
a common tell of ai generated powershell is a script that has dedicated functions to check types, often via several methods, and happily prints the output of the checks to shell. i do not get why it does this but it often adds dozens of lines that really serve no purpose but to make the shell output look fancy
It's frustrating when I catch it doing this because it's wasting tokens and time on what it reports as "one-off" scripts. Might add a requirement to never use shell scripting unless the task is truly a one liner.
Try being on linux and using zsh, it constantly fucks up scripts. Having it run commands via SSH into a windows machine is even more a comedy of errors.
I have to put in every claude.md that the shell is zsh. It's reached the point of annoyance I might just go back to bash.
> It's reached the point of annoyance I might just go back to bash.
You can use zsh as the default shell and still write your scripts for bash. Actually, it’s an advice I saw more than a couple of times. Otherwise you need to translate the bash-isms when you get bits of code from random places on the Internet.
Just put the right shebang (or ask Claude to do it). What’s the problem?
I find that Claude does this much more than other models. GPT, for instance, knows I'm on Windows running PowerShell, and writes very idiomatic PowerShell.
But maybe that's because I'm the sucker, and since PowerShell is more verbose it costs me more tokens than the terseness of Unix shells. Oh well.
Just run your llm in a container with the tools it needs. It makes your pc more secure by removing the risk of doing something destructive to your computer and the llm can do llm. Also, llms ssh'ing to the outside world? Asking for trouble later on
> you can use AI to understand something and map it to your own mental map
This "symbiosis" (for lack of better word) of human with AI seems to be an emergent value proposition of AI. In the process of doing stuff with AI, producing artefacts like code diffs, we are continuously able to decide how strong the mental map is of the current stage of the production process.
I could probably have worded this better but I'm sure it's something others have noticed... this choice we are able to make of how high fidelity our own understanding needs to be of the current working problem, and how that choice never really existed prior to AI.
> I could probably have worded this better but I'm sure it's something others have noticed... this choice we are able to make of how high fidelity our own understanding needs to be of the current working problem, and how that choice never really existed prior to AI.
I do get what you are saying and agree with you in principle, as I have noticed the same. But that said, this same way of thinking can really apply to any abstraction. It really just depends on the level you're working at. A software architect might know the nitty gritty details of how each service works, but really those implementation details don't matter if the abstractions are handled well enough, so in theory they don't really need to know those details as long as "the pieces fit". But of course the catch 22 there is you can't build good abstractions that can fit together well if you don't understand well enough the underlying details
> However AI cannot meaningfully handle feedback and learn.
Well this is the central bet of AI coding isn't it? We, the humans-in-the-loop, get better at knowing ahead of time which patterns AI will handle better than others, all the while the models actually get better.
reply