theptip

17 ms·

331.
by theptip 9mo ago
To be clear, Polanski doesn’t have authority to terminate anything, and this is a letter notifying Palantir that they will be cancelled _if the Greens win_. Quite misleading title in the article.
332.
by theptip 9mo ago
Yeah, it’s kind of meaningless to attribute this to AI without measuring the base rate. It’s for sure plausible that it’s increasing, but I’m certain this kind of thing happened with humans too.
333.
by theptip 9mo ago
This is mostly an ad for their product. But I bet you can get pretty good results with a Claude Code agent using a couple simple skills. Should be extremely easy for AI to successfully detect hallucinated references as they are semi-structu
334.
by theptip 9mo ago
For many reasons, you’re not gonna get into Anthropic with that attitude.
335.
by theptip 9mo ago
The point of TFA is that you are not letting it do whatever it wants, you are restricting it to just the subset of files and capabilities that you mount on the VM.
336.
by theptip 9mo ago
> Will create As others have noted, it’s a long-term trend - agree that as you note it’ll get worse. The Russian psy-ops campaigns from the Internet Research Agency during Trump #1 campaign being a notable entry, where for example they s
337.
by theptip 9mo ago
Hassabis put forth a nice taxonomy of innovation: interpolation, extrapolation, and paradigm shifts. AI is currently great at interpolation, and in some fields (like biology) there seems to be low-hanging fruit for this kind of connect-the-
338.
by theptip 9mo ago
Yes, of course, that is what I’m saying you need to change.
339.
by theptip 9mo ago
This is a sound tactical move to provide a hedge against future US economic threats. Strategically, I do think you want to be coming up with a plan to shield core industries like auto, shipping, energy, and some parts of manufacturing (eg “
340.
by theptip 9mo ago
It’s a good idea to have Claude write down the execution plan (including todos). Or you can use something like Linear / GH Issues to track the big items. Then small/tactical todos are what you track in session todos. This approach
341.
by theptip 9mo ago
Did you eval using screenshots or some sort of rendered visualization instead of the CLI? I wonder if Claude has better visual intelligence when viewing images (lots of these in its training set) rather than ascii schematics (probably very
342.
by theptip 9mo ago
Claude Code has had this feature for a few months now.
343.
by theptip 9mo ago
> Why does dev-owned testing look so compelling in theory, yet fall short in almost every real-world implementation Seems like a weird assertion. Plenty of startups do “dev owned testing” ie not hiring dedicated QA. Lots of big-tech does
344.
Demystifying Evals for AI Agents(anthropic.com)
2 points by theptip 9mo ago|0 comments
345.
by theptip 9mo ago
> When it became cheaper to publish text did the quality go up? Obviously, yes? Maybe not the median or even mean, but peak quality for sure. If you know where to look there are more high-quality takes available now than ever before. (An
346.
by theptip 9mo ago
Counterpoint - I think it’s going to become much easier for hobbyists and motivated small companies to make bigger projects. I expect to see more OSS, more competition, and eventually better quality-per-price (probably even better absolute
347.
by theptip 9mo ago
It’s a risk/convenience tradeoff. The biggest threat is Claude accidentally accesses and leaks your ssl keys, or gets prompt-hijacked to do the same. A simple sandbox fixes this. There are theoretical risks of Claude getting fully owne
348.
by theptip 10mo ago
No, they are not getting worse. Again, look at METR task times. The peak capability is very obviously, and objectively, increasing. The scaffolding you need to elicit top performance changes each generation. I feel it’s less scaffolding now
349.
by theptip 10mo ago
I’d say “skill issue” since this is a domain where there are actually plenty of ways to “hold it wrong” and lots of ink spilled on how to hold it better, and your phrasing connotes dismissal of user despair which is not my intent. (I’m dism
350.
by theptip 10mo ago
I do wonder how hard/expensive it would be to accompany every adversarial nuclear sub with your own underwater drone (perhaps itself nuclear powered) to track locations. Assuming you stay nearby after they leave port could you keep up?
351.
by theptip 10mo ago
Does it matter? Hypothetically if these pre-training datasets disappeared, you can distill from the smartest current model, or have them write textbooks.
352.
by theptip 10mo ago
They are not getting worse, they are getting better. You just haven't figured out the scaffolding required to elicit good performance from this generation. Unit tests would be a good place to start for the failure mode discussed. As ot
353.
by theptip 10mo ago
Integration with Function is a great use-case. There is a huge category of pre-diagnostic health questions (“Medicine 3.0” as Attia puts it) where personalization and detailed interpretation of results is important, yet insurance typically
354.
by theptip 10mo ago
Great point. And I think this is why I love the framing in the OP. “Do more” is a failure mode and path to burnout. “Do what I’m doing and you’re not doing” is a cue that an ambitious engineer can reflect on constantly.
355.
by theptip 10mo ago
Right - I’m missing how you get the source code in the OP. It says you tmux in with ssh agent forwarding for GH. But you can’t do that on your iOS device? So you have to set up all your repos in the morning before leaving the house, then co
356.
by theptip 10mo ago
Is this true? Nobody has a counter to ICBMs as far as I know. Defending against strategic bombers is a different game from submarine-launched cruise missiles / hypersonics too. If you don’t spot the submarine 50 miles off your coast yo
357.
by theptip 10mo ago
> Except that the use of nuclear bombs is predicated on the theory of strategic air bombing Aren’t ICBMs and submarine-launched warheads the other two parts of the US’s triad?
358.
by theptip 10mo ago
I dunno man. A useful psychological trick in this scenario is to imagine you were on the other side of the transaction. Say you have nine figures in your account. Would you push it all into this company to be able to have a fun job? Seems c
359.
by theptip 10mo ago
Things like https://github.com/textcortex/claude-code-sandbox seem like the bare minimum. There are a few other projects doing this. The first threat is making edits to arbitrary files, exfiltrating your SSL keys or cr
360.
by theptip 10mo ago
Guys, you obviously cannot suggest that —dangerously-skip-permissions is ok here, especially in the same paragraph as “even if you are not a software engineer”. This is untrusted text from the Internet, it surely contains examples of prompt

More ›