CityOfThrowaway

I have a feeling this person is using far-from-frontier models, totally disconnected from the development environment.

Using, like, gpt-4o is extremely not useful for programming. But using Claude Code in your actual repo is insanely useful.

Gotta use the right tool + model.

I read the post differently – the point of the exercise is not that you need to know the answers to the questions. It's to gauge your emotional reaction to the question itself.

By examining the types of tasks you will be consistently faced with, you can ask yourself, "Do I actually want to do that?"

Awesome!

The best practice that I've seen done (and enjoy) is:

1. Always squash and merge 2. But have tidy commits 3. So your squashed commit has a useful list of what is in the commit

I haven't tried this yet (though I plan to).

One thing I would love is if I could give it a hint and have it extract out certain types of changes into its own branch that could split into a new PR.

I often find myself adding a new, re-usable component or doing a small refactor in the middle of a project. When you're a few commits into a project and start doing side-quests, it's super annoying to untangle that work.

The options are one of:

1. A mega PR (which everybody hates) 2. Methodologically untangling the side quest post-hoc 3. Not doing it

In principle, the "right" thing to do would be to go checkout main, do the side quest, get it merged and then continue.

But that's annoying and I'd rather just jam through, have AI untangle it, and then stack the commits (ala Graphite).

It's easy to verbally explain what stuff is side-quest vs. main quest but it's super annoying to actually do the untangling.

Maybe this tool magically can do that... but I do wonder if some context hints from the dev would help / make it more effective.

I agree with this, in general. And I think having the base models improve their performance on being resilient against these types of attacks is a very good idea.

That said, my primary point was that the claims made in the paper are at best using the wrong terminology (called base models "agents") and at worst, drawing massively over-generalized conclusions on the basis of their own idiosyncratic engineering decisions.

They tested one specific agent implementation that they themselves made, and made sweeping claims about LLM agents.

Situationally, the original post claims that LLM Agents cannot do the tasks well. But they only tested one agent and swapped out models.

The conclusion here is that the very specific Agent that Salesforce built cannot do these tasks.

Which frankly, is not a very interesting conclusion.

I'm not sure that's right. You can write prompts that make use of, but don't disclose secret information.

I have valid live, production cases where we do this and don't have info-leaking problems due to the scaffolding / prompting techniques we use.

Part of the problem is that confidentiality is in the eye of the beholder so extra effort needs to be taken to make explicit what should and should not be shared.

That said, one valid conclusion that could be drawn from this research is that base models are currently insufficient at exercising nuanced judgment about what should and should not be disclosed without explicit instruction.

That's an interesting thing to know and would be a good place for model builders to put some effort.

The benchmark is useful, but the conclusion of the write-up is that current generation LLMs can't solve the problem. That's not a valid conclusion to draw. The results here tell us mostly about the skill of the agent-designer, not the capabilities of the model.

This paper doesn't make any sense. They are claiming LLMs are bad at this set of tasks, but the reality is that they built a bad agent.

I bet it's possible to nearly ace this using existing LLMs by designing a better agent. Better tool structure, better scaffolding, better prompting.

LLMs are not gods, they are tools that require good engineering to achieve good outcomes.

#2 and #3 are both just engineering problems at this point.

The foundation models don't adapt quickly, but you can definitely build systems to inject context that changes behaviors

And if you build that system intentionally and correctly, then it's handled for all patients. With human doctors, each individual doctor has to be fed context and change their behavior based on the information, which is stochastic to say the least.

I mean... Taylor swift is literally the biggest musician in history... right now...

There's still enormously mainstream culture. Even more enormous in the fat tail than before.

There's just also a shocking depth in the mid tail now too.

The problem with movies is that Hollywood killed itself and tech helped. Movies and TV just suck now, for the most part.

Music, fashion, and visual culture are still alive and well.

You should make a page with some browsable pre-generated pages and post again to spark interest

All 501(c)(3) are funded in part by the public by way of uncollected tax revenues for economically valuable activity.

I was talking generally, not this specific set of cuts. From what I know (and what is generally accepted), there is an abundance of waste in government spending. I don't know anything about this specific Accenture contract, so can't say much about it. But I have enough close contact with government to know with absolute certainty that some amount of cleanup is long past due.

The data on hostile takeovers by "corporate raiders" very much does not support the characterization here.

The category of PE firms you're talking about buy companies that are deeply troubled. Generally due to the management's unwillingness to accept reality and make change, the company is heading towards oblivion one way or another.

Perhaps surprisingly, the vast majority of takeover targets wind up as net job creators on a 5-10 year time horizons. That's despite the fact that they do usually start by divesting assets that don't make sense and laying off non-productive employees. But divested assets aren't generally killed – they are usually sold to somebody else who often does something better with it.

Also, companies conduct massive downsizing and rationalization all the time when in distress, and not only when they are taken over by a "corporate raider".

In the private markets, these actors are definitely distasteful. They do cleanup work that feels bad, and they often get rich doing it. But they also serve a necessary role in the markets.

Companies that are egregiously misusing capital and resources are a drag on the economy. It's a bad thing for there to be a bunch of zombie companies holding onto assets that could be used in better ways.

A more generous framing would be something like a home flipper. They buy properties that are a mess, clean it up real good, throw out the old stuff for recycling, install some modern appliances, and sell it to somebody else.

One of my laments is that there is no automatic equivalent force in the government. Agencies grow and grow, projects grow and grow, all totally decoupled from whether they are achieving any progress whatsoever towards the agency's mission.

I'm not defending the specific actions of this administration (for which I simply don't know enough), but it is refreshing to see the government rummaging through its mess and cutting stuff that is irrational, corrupt, and not serving the mission.

Another thing I really like about Kasia's milk system is that it is a zero-emotional overhead communication. Just take the ticket and put it on her desk, and all will be handled.

He sets up a definition where "real intelligence" requires consciousness, then argues AI lacks consciousness, therefore AI lacks real intelligence. This is somewhat circular.

The argument that consciousness can't be computable seems like a stretch as well.

Perhaps not widely known – there are two autonomous regions in Myanmar governed by the "Wa State" which is not recognized by the international community but is very much in control of their territory.

It is financed almost entirely by crime. First through growing poppy that became heroin, then through enormous meth labs, and now also through crypto and digital scams.

The GDP of the autonomous region exceeds the GDP of the recognized Myanmar territory, despite having something like 10% of the population.

Also of interest – the ethnic subgroup that controls the region were literal head hunters up until the mid 20th century! Through a combination of CIA operations, evangelical Christianity, and the CCP (all in conflicting power struggles), the Wa people swore off head hunting and "modernized".

Anyways, there's a very interesting book about the history here called Narcotopia [1]for those interested. It uses as a primary source one of the key actors in the modernization, and as a result is incredibly biased. But super interesting nevertheless.

[0] https://en.wikipedia.org/wiki/Wa_State?wprov=sfti1 [1] https://a.co/d/bYkcrNg

Of course everything is renegotiable. But after reneging on every serious promise you've ever made, I do think it's worthwhile pointing out that this organization is shamelessly Machiavellian. They have mistreated all of their constituencies pretty rapidly!

Yeah... this is consistent with my experience. Try to keep calm and actually try again in a few hours.

It sounds like you have 2FA enabled so an account takeover is very unlikely. It sounds like you haven't violated the TOS in some egregious way, so actual ban is unlikely.

Most likely is there's some type of crazy cookie thing going on that is causing the auth system to vomit. It's shocking that this happens, but alas.

OpenAI really does seem to embody the idea that all contracts are renegotiable if you're Machiavellian enough.

YMMV but I've had similar loops during login that were resolved by trying to login in incognito mode. And if it works, then clear all your cookies / reboot / try again in the main browser.

There's something broken in google's auth security system.

This feels dangerously close to being almost totally automatable, unfortunately

Even as a very experienced engineer, I find that I learn a ton from chatting with Claude about technical decisions.

It is so valuable to take the ideas and questions in your head and get them out. You can then have Claude interrogate and problem solve with you.

By default, I think Claude is very conservative and sometimes comes up with ideas that aren't quite right. But you can actually just tell it to be less concerned about X or Y, to start over from scratch, etc.

Another trick I use is to ask it to ask me one question at a time to help clarify my thinking. I find that the one by one questioning really helps me find the boundaries of my knowledge and crystallize my opinions.

Yes, but if they did their own language it would have almost certainly died in obscurity.

React won because it solved a real pain point for a group of people, using the techniques they were already using... but better!

Having built large apps in angular, handlebars, and jquery... it was (and still is) a godsend for building highly interactive, stateful web applications.

I am deeply skeptical of some of the stuff that people are doing now wrt. putting backend logic into React components, but the fundamentals of React are still very reasonable to me.

The global grid concept is a fascinating piece of infrastructure engineering.

But the anti-market rhetoric here feels completely backwards. The article cites Shell building Europe's largest renewable hydrogen plant, BP's $30B Australian investment, and major moves by other energy companies... then somehow spins this as evidence that corporate involvement is problematic?

Market incentives and existing energy infrastructure expertise are exactly what's driving the current clean energy buildout at scale.

I wish this was presented as the interesting systems engineering challenge it is, rather than wrapped in reflexive anti-corporate ideology. The technical aspects deserve serious analysis without the political baggage.

Blacklist is too concrete.

With some domains, you merely will find a higher % of your emails land in spam, or your content ranks a bit worse, etc.

There's a somewhat random continuum. Haunting is a funny word that does sort of include some variability.

For readers: you could build Namefax as a startup! Pure-partnerships based model... distribute it through registrars.

"Heads up, this is a pre-owned domain. Do you want to get the Namefax for $0.99 before you buy?"