I think it is an interesting finding and it would be interesting to hear more about the implications from someone in the field.

I only broadly understand this is some kind of peeking behind the curtain of some process that was not fully understood before

My most successful autonomous runs have been expanding scrapers across a number of similar but different portals. I had examples and targets and it just kept searching for new ones and adding them.

But even doing basic ML auto research k have found it to be surprisingly poor except at trivial but useful augmentation of models. Yes it can implement things but somehow I am required a lot even though I set up a lot of framework around it.

My mental model is that it's very good at complex deterministic work like reading bad API docs and getting some connectors to work.

But perhaps I care less about being stuck in a local optimum there.

When I was working more vision was always a bottleneck ... Staring at yet more close things would be less useful than staring at far away things

If this was some kind of crime they would have censored information for a long while if not clearly correct.

This is the problem.

I think there is a huge gap between people on salaries getting effectively more responsibility by being given spend that they otherwise would not have had and people hustling on projects on their own.

Yes it is 100% what I use but I am never happy with usage. It burns up by sub fast and there is little feelings of control. Experiments like using lower tier models are hard to understand in reality. Graphify might work or it might not. I have no idea.

I have even had founder level emails that presumably are confidential sent to me because I share the name of someone operating in tech.

I respond or report when it's obviously some real person running a small group but for large monoliths there is very little to do except quickly reply to corporate email.

Really wish there was some kind of high level discussion about building something for this specific problem of non malicious wrong person same name errors.

Google could do it it's just not something that is monetizable at a scale they care about IMO and I have not been able to think of a way to make this work operating outside of email monoliths.

Would love to hear if anyone has ideas.

Collective notions of shame, actual networks of friends and families that reinforce correct behaviour or issue corrections.

Think about simply how credit networks form and function. And why visiting a food truck or medieval travelling doctor for your vial of ointment is different from buying special products from a brick and mortar establishment.

Basically if you or the network has a harder time back propagating defaults and bad credit in a way that prevents future bad outcomes then that is a loss of high trust.

This isn't about race really unless you are operating at the level of some biological or genetic connection to behaviour ... But that is a pretty strange place to be as there a whole host of confounding factors that are much more obvious and believable and I cast serious doubt that even a motivated racist would ever credibly be able to do empirical studies showing causal links between any given genetic population cluster and the emergent societal behaviour. These are such high dimensional systems it just seems insane to even think one could measure this effect.

The invisible substrate is the society unfortunately ... And we are all bad at writing it down and measuring it.

There are probably three modes of safety.

Deploy tech to unknown group. Could be enemy could be friend. You disable it's abilities perhaps.

If you deploy tech to friends you might enable more defense.

Anthropic models seems to have unstable safety predicates that have a hard time advising on situations regarding preservation of a people.

The huge problem is that humans AND ai both seem to fail at understanding how humans are made and which human are which.

You are uniquely responsible for protecting your people. You can not simply funge their people for your people and pretend that is a fine trade off. And even beyond that these safety predicates appear to not have any notion baked into them of diversity or TFR or lineage. The models view the descendant of a nearly extent lineage the same way they view the descendant of a high TFR lineage.

You can have ANY kind of opinion on this but this naive no opinion vague word based safety predicates is very scary and dangerous.

I am deeply worried about Anthropic as I have yet to hear anything that makes me think they have real adults in the room. I would love to be wrong and so I write here. Please do let me know if there are good things they have written on this.

Do they have even a basic understanding of the different regimes of safety and what allegiance means to ones own state?

It would be fine to say they are opting out of all forms of protection against adversaries.

But it feels like just more insane naive tech bro stuff.

As someone outside the tech bro bubble in fintech in London, can somebody explain this in a way that doesn't indicate these are sort of kids in a playground who think there is no such thing as the wolf?

Again, opting out and specializing in tech that you are going to provide to your enemies AND friends alike is fine. That is a good specialization. But this is not what I hear. I hear protest songs not deep thinking of thousand year mind set.

I thought during the pandemic FB marketplace was going to go somewhere. I thought they would try to solve physical delivery with like an Uber service and credit network for financials etc. it would be huge.

But no. What has happened is that primary dealers are now flooding marketplace with fake low ball posts to make it unusable and destroy the secondary market.

I recently was shopping for bunk beds and lo and behind there were hundreds of not thousands of posts just for my local area all from maybe a dozen or so accounts created around 2023.

This is somebody's business (spam order flow as a service) and I assume that they pay fb enough for some API they fb literally doesn't care.

My theory is that every single feature on FB is a/b tested to be as bad as it can be if it maximizes screen time. Search doesn't work. You can't find your profile settings or feeds easily. All on purpose to maximize the time you spend there.

The feed has been dead for me for ages. I would recommend many users simply use it as a storage log book and increase FB costs by requesting all your data occasionally.

It's one of the worst companies out there for explicit bad behaviour IMO.

Pre applying is usually less stressful as you are trying to do your best to show that are ok. Just showing up is a not great for either side.

I often wonder the reasoning.

I guess we could call this gathering and then do your best conditional on what you found right now.

It explains what it means.

It's weird, it's like many agents are now in a phase of constantly getting more information and never just thinking with what they've got.

So they build useful things and then make them pretty bad and less useful. If they were useful your interest or need would complete and you would move on.

Fundamentally I think it is important to say this. Addiction confounds some things in the space of designed systems

i.e. imagine a change that is literally a small diff, that is easy to describe as a mere user and not a developer, and that requires quite a lot of deep understanding merely to submit as a PR (build the project! run the tests! write the template for the PR!).

Really a lot of this stuff ends up being a kind of failure mode of various projects that we all fall into at some point where "config" is in the code and what could be a simple change and test required a lot of friction.

Obviously not all submissions are going to be like this but I think I've tried a few little ones like that where I would normally just leave whatever annoyance I have alone but think "hey maybe it's 10 min faff with AI and a PR".

The structure of the project incentives kind of creates this. Increasing cost to contribution is a valid strategy of course, but from a holistic project point of view it is not always a good one especially assuming you are not dealing with adversarial contributors but only slightly incompetent ones.

If you look at tool calls like MCP and what not you can see it gets ridiculous. Even though it's small for example calling pal MCP from the prompt is still burning tokens afaik. This is "nobody's" fault in this case really but you can see how the incentives are and we all need to think how to make this entire space more usable.