I found it worthwhile reading how auto mode works[0]. The short version is that the classifier only sees user messages and the tool call. So if your agents communicate via tool calls, the classifier doesn’t see that text at all (or any other prompt injection) and will block a potentially dangerous command.

[0] https://www.anthropic.com/research/claude-code-auto-mode

auto mode is a bad joke

You can access Claude models through any number of providers - OpenRouter, AWS Bedrock, etc. and as far as I know, Claude Code can be used with any of them. The only real benefit to going direct through Anthropic is the subscriptions. It’s worth the risk for the discount, but if you get locked out, there’s other (pricy) options to get unblocked.

I don’t know about you, but for me materially useful AI costs around $10,000/mo in tokens. But it is very useful.

Claude subscriptions help, but only if I come in with the expertise to build the entire system around the model and I don’t mind the whole thing grinding to a halt when my session limit is saturated. Anyone selling the system off the shelf is going to be charging API rates. Computers were cheap in comparison.

I can’t speak for everyone, but I would intuitively venture to guess there would be some greater than marginal productivity benefits from say, a $200/m Claude sub for each employee in many organizations outside of tech.

Perhaps they’re all on the $20/m plan which is useless?

Maybe, even likely. But how do you observe this? We are in this situation at work, it feels like "sure AI is helpful" but do we actually deliver any faster?

“Marginal” improvement isn’t really going to be enough. Products need to be closer to 10x better than the alternative to get traction (which can be divided across several dimensions). AI, I think, is just starting to get there.

I get what you’re saying but I’m just not a fan of this type of argument. It always paints AI as this “in the future” thing which absolves any current criticisms or observations.

LLMs are great and definitely make you more productive, for some reason 493/500 firms haven’t been able to figure that out even though it’s trivially easy to provide employees with frontier models and it’s also trivially easy to use frontier models.

Which is why I say comparing it to personal computers in 1987 is not an apt comparison. That took redefining physical workspaces, new infrastructure, and new personnel which wasn’t easy to come by. Becoming a company that uses AI is just as easy (if not easier) as using any other SaaS.

See sibling comment. 10x measured how? In pure coding tasks under certain conditions the speedup is huge, but seems really unclear how this would transfer to any significant speedup across the whole undertaking (project, delivery etc).

Or just in enterprises where they can't have plans.

Holy shit. I’m spending 2 orders of magnitude less per month and still getting good value. But, I’m not an AI maximalist.

If it costs 2x-3x the median household income to be an AI maximalist that definitely puts some comments on this forum into context.

It’s a business-scale cost. When I say materially useful, I mean replacing multiple salaries and improving quality at the same time. Not some variant of “make an existing employee 20% faster.”

The personal hobby version is to get the Claude max plan.

I came across this recently and was initially surprised too. It makes sense when you consider that any company trying to build models (or an agent/value-add layer on top of a model) needs benchmarks. The best funded will do the full vertical - build their benchmarks and models together - but many will outsource the benchmark. This is particularly valuable when the dataset is difficult to compile (robotics being a prime example) - it makes more sense to let the benchmark vendor focus on that.

Sure it’s useful, but I’m asking who actually pays for this. Sort of a conflict of interest if you’re providing public testing for the companies paying you for the tests no?

Perhaps it’s purely for VC so they can judge which model is best? Also seems a little too conflict of interest-y, but that might just be too cynical.

Wouldn’t it just be easier to ask their team to upvote the post? They probably all have legitimate hacker news accounts already and are happy to upvote their work.

Not really.

It’s more of a fun anecdote, but people forget that there was once a plague of stolen Apple devices. Apple customers were frequently targeted by pickpockets, etc. because of the value of the devices. I have no evidence but I’ve always believed this was the root of the privacy focus - Apple could lock devices to biometrics and eliminate various workarounds, making stolen devices worthless, and the thefts evaporated practically overnight.

Phone theft is a growing problem again, but it’s still less common than it was because they need to get an unlocked device, and I don’t hear them targeting Apple in particular.

> when you want to let them jump in and out of multiple apps (e.g. pick what game they want to play on their own)

The solution to that is Assistive Access.

It’s more painful to use than it should be, but it absolutely solves this problem.

I’ve had the privilege to work at a company with a highly aligned legal team. It turns out there are quite a few people who get law degrees and are just sensible, chill, fun people. And if you can hire them onto staff, there’s no limit to the awesome things you can do.

I've never frowned while paying a legal bill. They do good work, provide useful perspective, avoid problems before they start... reminds me of the better engineers I know, but for contracts and compliance.

Which also seems a bit silly because rural areas have plenty of events and many use e-tickets. Maybe OP just doesn't like events.

GPUs aren’t becoming outdated in practice. An H100 still rents for something like $3/hour. The only reason to take them offline today would be due to limited data centre capacity. In space, that doesn’t apply.

For now, at least, this market doesn’t behave like the one for CPUs which are more fungible and general purpose.

Rather than thinking about the value of a GPU over time, it’s more useful to think about the value of a model (or generation of models) that will run on the GPU.

Sol 6 launched last week. People will build products/agents/systems on Sol 6. Many of those will end up staying on that model for the lifetime of the end product, which means there will be long-term demand for the model, and that means long-term demand for the GPU set that the model runs on.

For now, at least, models are tuned to a specific GPU and don’t move to the newest hardware very easily.

At the end of the day, will the lifetime value of a GPU exceed the cost of launching it into space? That’s TBD. But there’s growing evidence it might.

> In space, that doesn’t apply.

How does capacity not apply in space? You can’t just boost a gpu into orbit and it’ll start running. It needs to be hooked into power and cooling infrastructure still and that will always be limited by how much infrastructure you have up there.