idopmstuff

"The study, authored by Capraro with Chiara Marcoccia of École Normale Supérieure and Walter Quattrociocchi of Sapienza University of Rome, deliberately used questions where AI models typically fail: visual details from films, such as the colour of a team’s uniform in Bend It Like Beckham."

I get why they used questions where AI models fail, but it also really reduces the value of this study. Nobody is really asking AI the color of a team's uniform in a movie, and if they do and confidently get it wrong, it just doesn't matter at all.

Asking trivial questions also feels like it would affect the rate at which people are willing to confidently say things that are wrong. If you ask me some question of pointless trivia and I ask ChatGPT, I'll probably just repeat the answer because who cares. If you ask me something even mildly important and I ask ChatGPT, I'll either verify the information before I repeat it to you, or I'll qualify that I looked it up with ChatGPT and didn't verify. But some things are just so unimportant that they don't even warrant the disclaimer.

Yeah good god OP please start writing a Substack about this! Engineer who buys a run-down bowling alley and cleverly fixes problems is exactly what the internet needs more of.

Ah, that is a fantastic idea, it's on my to do list.

Thanks! I am definitely not an expert in either, and I've run the content by both Fable and GPT-5.6 with instructions to make sure it's written in such a way that it would read normally to people in the industry. They assure me the wording makes sense in that context, but we'll see if the sites actually get traction or not.

I heard an episode of the Odd Lots podcast about HayWire (haywireag.com), a site that pulls public data from government PDFs + APIs, uses LLMs to parse it and turns it into an easily readable website that has all of the latest info on hay prices.

The host made an offhand mention that there's probably a bunch of other similar sites that could be created with all the of useful but difficult-to-access government data out there. That sounded interesting, so I thought I'd give it a whirl!

Working on a few of them, including The Waterline (https://the-waterline.com/) for water info for the western US, The Scramble (https://the-scramble.com/) for egg prices, and The Dwell (https://the-dwell.com/) for container ship dwell times.

All pretty fascinating topics to learn about, plus it's been interesting to see how much of the website setup I can fully delegate to Claude. With Cloudflare to buy domains and put the sites up, a Google Service Account with access to Google Search Console and GA4 to create those properties and a Buttondown API key for weekly email sending, it's almost all hands off for me. Though it refuses to take control of the browser and create a new Buttondown account, which I was surprised is a red line.

I assume you're referring to this:

And when the brain encounters something it can’t process efficiently, it doesn’t simply adapt. Brain imaging studies cited in the review show it generates stronger neural responses in visual areas, consumes more oxygen, and in some people produces pain, distortion, or worse.

If the studies are of a person's initial exposure to these sorts of conditions, then that doesn't tell us anything about whether people adapt over time (and to be clear I have not read all the studies, but given the limitations listed I'm comfortable assuming they're not incredibly robust until someone tells me otherwise). I suspect the article's use of the word "adapt" is not the same as mine; from the context when they say the brain doesn't adapt they just mean that it shows a response at the time of the particular exposure they're measuring.

The Limitations section at the bottom certainly has a lot of limitations:

This paper is a review, meaning it synthesizes and interprets existing research rather than presenting new experimental data. The authors themselves note that current visual tests for susceptibility to discomfort are subjective and poorly standardized. They also acknowledge that the proposed mechanism (that discomfort is the brain’s response to overwork) has not been fully tested, particularly the hypothesis that colored tints reduce discomfort by steering visual stimulation away from overactive brain areas. The relationship between the brain’s excitatory and inhibitory chemical signals and visual discomfort also remains, in their words, “unsettled.” Several key research questions are flagged as unresolved, including how to best quantify the real-world impact of visual stress on people’s lives and how to objectively measure susceptibility.

Flickering lights are about the only thing I saw in here that seem like they'd be a problem in the long term. Everything else your brain just adjusts to over time and stops noticing. Maybe the first few days in an office with bright colors would be slightly distracting, but after that you just stop seeing them. I would guess that a lot of the studies they reviewed probably tested people's reactions to these things when they saw them one time, not the hundredth time.

I own a dozen Amazon brands that are probably largely of the kind that OP would want this to get rid of (sourced from China, not name brands by any stretch, only sell on Amazon). For the most part I would say this extension is not a great idea (obviously very biased!), since I purchase brands that have high quality products that typically have pretty poor branding/online presence that I can improve. My stuff is very frequently of the same quality (and sometimes from the same factories) as much pricier stuff but at a lower cost. To some of those suggesting you can get this stuff on Aliexpress, in some cases that is true, though of course the big benefit of buying from Amazon is that there's no risk to buying no-name stuff because if it's junk you can return it.

In any case, I gave this a try to see which of my brands it would filter out. It's weirdly inconsistent.

One of my brands was filtered out because there's no brand name at the beginning of the listing. That's just an outright bad rule, because Amazon generally decides whether or not the brand name appears first. This brand is trademarked and has Brand Registry, so it qualifies for that treatment, just not getting it right now. Also, a number of other brands without brand names did not get the same treatment (and these are very much the type of products this is designed to filter out).

On another one, it misunderstood the product model, which is at the beginning of the product name, as the brand and hid it based on that. That one was a bad one because the model is only three characters, which is extremely unlikely for a brand name.

One product I sell is a hunting accessory, so I did some searches there. It hid everything by the brand KUIU, which is a well-known and very high end hunting brand. Definitely wrong there.

So yeah, sort of an interesting idea, but the execution is pretty sloppy and the creator clearly doesn't have a full understanding of how Amazon listings work.

It might seem obvious to coders, but the difference between Claude Code and Claude.ai's chat is enormous, even if those two run the same model.

In my experience, Claude Code is vastly better for doing tasks, writing code, etc., but Claude.ai is better for analysis and high-level planning. When I'm working on a new project, I've started using the latter to do the initial planning, get feedback and draw up a spec, which then goes to Claude Code.

For this project, I probably would've done something similar - use CC to get whatever you need out of the image files, but have Claude.ai do the actual review/diagnosing.

Either way, I often think about how far behind most of the world is in really understanding AI. The overwhelming majority of people would never guess that you get vastly different outcomes from the exact same model in a different harness (tbf most people don't know what a harness is). I spend hours every day using AI for a broad range of tasks and still feel like I know a fraction of what there is to know. I haven't even tried the new GLM model (or really any of the open source Chinese ones of the most recent generation). With so many people thinking that the free version of ChatGPT is SOTA AI, a lot of folks are in for a very rude awakening at some point soon.

How is AI involved in this? None of the features seem like they would need to use it at all. So presumably AI means it's just vibe coded?

Also, the comparison chart is just dishonest. People didn't e-sign proposals before your product? Come on.

Not worthwhile feels a bit strong (a good FAQ is definitely worthwhile!), but I definitely agree that there is a big difference between any kind of art (writing, playing music, creation of images/videos/etc.) for its own sake and for commercial purposes. AI is terrible for the former but perfectly fine for the latter.

There will always be value in a human writing fiction or a memoir or even a Substack. The human perspective is inherently valuable there. Much less so with ad copy that's just going to get A/B tested ad infinitum until a winner is picked out based entirely on data.

Same with visual art. Art painters aren't going to lose their jobs to AI, but once you've got a robot that can paint a house reliably, house painters are done for.

To investigate whether skills are being lost in the field of computer science, researchers at the AI firm Anthropic in San Francisco, California, designed a randomized controlled trial in which 52 software engineers were asked to perform a basic coding task3. During the exercise, all 52 participants could search the web and access instructions on how to do the task. Half of the participants were prompted to use an AI assistant as well.

Afterwards, all of the software engineers were asked to complete a quiz about what they had learnt from the task. The participants who had used an AI assistant did significantly worse on the quiz than those who hadn’t: the average score was 50% in the AI group versus 67% in the non-AI group.

This doesn't strike me as a great test? Most engineers aren't going to learn anything from a basic coding task anyway, so I do wonder exactly what they were testing there. If it was just recall about what the issue was, then it doesn't really strike me as a problem - using AI to handle simple problems that it's clearly capable of dealing with is the right way to use it, and of course you're not going to spend time poring over the details because then you haven't saved any time by using AI.

There are other examples that don't strike me as particularly problematic, like GPS eroding people's sense of direction. It's totally reasonable to let a skill atrophy that you no longer really need because you have an ever-present tool to handle it. I'm a lot worse at doing long division than I was when I was <whatever grade one learns long division in>.

The whole skill atrophy thing seems like much less of a problem than it's made out to be. We've been letting skills atrophy for good reason long before the advent of AI. If you start at McDonald's as a fry cook and work your way up to regional manager, if you suddenly have to work a shift on the fry station you're going to be worse than you were when you were doing it all the time. MDs at investment banks almost certainly can't put together a pitch deck as well as the junior bankers who are doing that task regularly. These things are fine - part of moving up in the world and having a broader impact is being able to successfully delegate tasks, and when you delegate tasks your skill at those tasks will atrophy. No real difference whether you're delegating them to AI or not.

To be clear, there are of course cases where skill atrophy is bad. iLoveOncall posted about senior engineers in their org who have lost all of those skills and their judgment along with them. That's definitely bad! If you delegate so much that you lose the ability to even judge good work, now you can't even delegate effectively any more.

I think the real lesson with AI is that you need to be self-aware about what skills you should practice and retain vs. what skills you can let atrophy, since it's easier than ever to hand things off. I've lost most of my ability to write a SQL query, but that's fine because it was only a skill I used intermittently and AI can always do the job fine at the level of complexity I need. I have not let my skill of writing product specs atrophy (I am a PM, in case you haven't read my username), because that's critical to using AI correctly in the first place.

So many! I manage a fund that buys small e-comm brands, and at this point the whole thing runs on a combination of AI and tools created with AI. My favorite is one that scrapes my Alibaba/WeChat/WhatsApp/email supplier convos daily and uses that to build a dashboard tracking the status of my orders.

I write a Substack about the whole thing and have a pretty comprehensive list here: https://theautomatedoperator.substack.com/p/15-ways-im-using...

For the most part if you give it sample photos that have sufficient coverage of different angles, it's very good at faithfully reconstructing the product from whatever angle you choose.

The one exception I've encountered is baby mobiles. It really does not understand the physics there.

This is definitely the most interesting question in a ton of AI applications. I think folks should be really be spending a lot of time on figuring out how to deterministically check AI outputs in a way that's reliable in order to reduce the amount of work a human has to check, and to build tools that speed up the checking process.

Thinking about all of the fake citations in legal submissions that have come up of late, it seems pretty straightforward to set up a regex that captures all forms in which a cited case might be written (I could be wrong but I'd assume there's some standard variety of formats) and search those against a database (again assuming such a database exists) to ensure they all exist.

Then for the tougher problem of making sure that the cited cases say whatever the document citing them says they do, you could have an LLM run through the document, pull out the text with the case name and text about why it's being cited, then read the case and try to determine whether the reason for citing it is valid. Rather than just give a yes/no, you'd put the doc in front of the user and let them jump from citation to citation. On each citation, it'd pop up a card that shows the literal text of why it's being cited, a judgement from the LLM of whether it matches what the case says, and snippets of text from the case as evidence + deeplinks to that text within the case.

Or maybe you wouldn't even want to give the LLM's judgement since people might rely on that without reading, but there's definitely a way to speed up the review.

I believe OpenEvidence does something like this with medical papers. If you ask it a medical question, it doesn't answer so much as link you directly to the relevant papers so you can read them and determine if they're useful. Avoids all of the potential risks of using an LLM but still hugely valuable and time-saving for docs.

Two of them:

1. ChatGPT 3.5 wrote me a script to pull some data out of Shopify and write it to a Google Sheet. Nothing remotely impressive by today's standards, but I had just commanded a computer to write code in plain English and it worked!

2. I own a bunch of e-comm brands, and with every new image model I tried to get product photography. Nothing worked until Nano Banana Pro, when suddenly I gave it a crappy iPhone pic of a product and got back a fully usable whitebox photo of it. Then I tried making the sort of infographic-style images you usually see on Amazon, and it nailed those too! In hindsight they weren't perfect, but more than good enough to use. I was about to ship that product to my photographer, and I would've had my designer make the infographic images, so that was the first time AI actually replaced a human contractor for me. Pretty big "Oh shit this is going to seriously impact employment" moment. Wrote about it here: https://theautomatedoperator.substack.com/p/ai-just-took-my-...

Communism's failures are due to human social factors, not technological. You can't fix social problems with technology.

Yeah, that's fair. The argument would be that the core reason communism failed is because people inherently want to have greater status than those around them, so the ones who were in charge used their power to grant themselves a greater share of resources in order to demonstrate their greater status. If we have infinite non-scarce goods, whoever's in charge can still let everyone have as they need of those while demonstrating their greater status through non-scarce goods.

To be clear this isn't a prediction (I have no idea what's going to happen!), just the case I could see for this being the first version of communism that works. Though also it's not really communism, because everyone is not in fact equal; it's something like a pseudo-communist giant welfare state.

Oh yeah, to be clear I fully agree with everything you've said. My core argument is that there will be sufficient economic productivity for everyone to live incredibly well. Whether or not that happens depends entirely on the people who control both economic and political power. That keeps me up at night, and things could go horribly wrong.

I guess the optimistic side of me thinks that benevolence wins out because there's no cost to it. There is plenty of competition among the wealth for scarce resources, but food, medicine, and mass-produceable luxury goods are effectively free. Given that, it's probably just easier to give those away to everyone than to crush most of humanity by force. But that is absolutely naive optimism, because I really have no control in this situation and prefer feeling naive optimism to pessimism.

And on the communism front, I will just say that I find it some combination of deeply amusing/ironic/depressing that the people on the far left protesting AI because it'll take jobs are protesting the very technology that could, in fact, lead to the first successful incarnation of communism!

Sure, but the question is at what layer of abstraction do you have to prompt the AI?

You used to have to prompt the AI by starting to write the actual line of code you want, which it could autocomplete. Then you had to prompt it to write simple scripts or functions. The amount of scope you can prompt keeps getting bigger and bigger. Eventually, you have a PM or a CEO just telling it what features you need. Maybe it's a PM and a designer and a CEO and a CTO, but it will eventually get to the point where the number of people you need to do the prompting shrinks orders of magnitude from company sizes today. Maybe you just give the AI some money, prompt it to start a money-making business, and it goes out and does the same research and analysis that a seasoned entrepreneur would do to find an opportunity then builds out the business from there.

the results won't be any good

Maybe, but I wouldn't bet on that. The trend over time has been that the results from prompting AI to do things have gotten better. I used to prompt it to build me dashboards and it would fail spectacularly. Now it one-shots them. Maybe the code is terrible (though doesn't matter for me, I'm the only one using it and I can verify the dashboard content is correct), but if the trend continues, it'll get better. Maybe the trend won't continue, but I've yet to come across a good explanation of why AI capabilities will just top out and cease improving forever.

Even if we start building a lot more stuff, you still don't need those construction workers. You have a GC to manage the whole project, aided by an AI who's handling scheduling/operations/logistics. You have a detailed plan to build against.

So why do you need former construction workers to manage the robots? Why can't the GC and management AI run the whole thing?

Maybe there's some scenario where you still need something like one licensed person from each skilled trade to be responsible for the robots employing those trades. But there's no way you need everyone who worked on building sites managing robots, no matter how much construction you're doing.

Why is this time different?

That depends if AI gets to the point where it can fully replace workers, as opposed to just augmenting them. I heard Alex Imas on a podcast recently talking about how a SWE can be running 10 agents to be 10x as productive, then that SWE is more valuable so firms should want to hire more SWEs and pay them more.

That works for a while, but what if AI gets to the point where it can manage the 10 agents as well as the SWE? Of course you could say the SWE can now manage 10 agents who each manage 10 agents so he's even more valuable, but that has to break down eventually. You don't need 1,000 SWEs each managing 10,000 agents - you hit a bottleneck in the ability to give them work fast enough (even if you need the SWE at the top at all).

I think it's easier to think of from the perspective of blue collar labor. It's further out there time-wise, but let's assume we get a humanoid robot that can do any labor a human can do. It costs $25,000 and maybe a couple grand a year to operate. Works 20 hours a day when it's not charging.

The construction worker it replaces isn't going to start managing a team of robots construction robots - there's already a GC doing that, and you can't scale building nearly the way you can scale writing code because of physical constraints. When the robot I've described exists, a huge swathe of the population is going to be unemployed. There's no competitor to hire them because the competitors just get robots too.

The problem with the studies is that they're cases in which specific groups within the broader economy lost jobs. Those aren't really comparable to the (theoretical) path of job displacement of AI for a couple of reasons:

1. Those people didn't get substantial, ongoing financial assistance. If we end up in a UBI world, particular one where the UBI people get is high enough to get more or less anything that's not very scarce (e.g. land in coastal cities), the negative economic component of job loss is removed. 2. Everyone else still had jobs. When you lose your job and everyone else continues to work and be successful (or at least you perceive that to be the case), there's a big hit to the status and meaning in your life. If everyone is affected in the same way, then your relative status to others remains unchanged, and everyone collective needs to reorient society to find their meaning.

I'm not saying it will go well, but I do think there's a theoretically possible path where there is large scale unemployment but because we have nigh infinite productivity, everyone has access to unlimited non-scarce resources (including luxury cars and fine foods and whatever medical treatment they need), and we end up with an enormous number of competitive leagues of everything, events centered around music and arts, dinner parties and all manner of other social activities that are what give people meaning.

My feeling is it's not as bad of a metric as people think. Companies don't fully know the best way to use AI and things are changing rapidly, so you want people using a lot of tokens even on stuff that seems maybe kind of dumb on the surface, because if you find one useful thing and share it in the org that makes up for a lot of failures.

But I do think you also need to say, "To be clear, don't game the system. Any token usage that is even remotely justifiable as useful for the business is fine, and we will give you a lot of latitude. But if you're in the top 10% of token users, we are going to review your token usage, and if we find that you have a dozen agents perpetually running writing slam poetry, you're going to get fired."

Great stuff. My favorite genre of writing about AI is seeing how it can be practically applied to non-tech jobs/businesses. Wish we had more of this.

I'm curious about the 60% automation of financial/forensic analysis - what's missing? Is it stuff that's purely blocked by model capabilities, or are there places where scaffolding is likely to bridge the gaps?

Also curious about the workflow - is this more individual, LLM-driven features or agentic workflows? Looked like the former from the product video but there wasn't a ton of UX shown there.

I ask largely because this seems like the sort of thing where you could really start to string these features together in such a way that you start with a description of the case and whatever files you have, and then an agent does its analysis of the docs, spins up action items (get missing docs, confirm that X ambiguous doc is what the AI characterized it as, etc.) and tracks the progress of all of them, leaving your forensic accountant there in a supervisory role, managing and providing expertise.

It feels like that's the way a lot of expert analysis jobs like this are headed. I've been working on the same sort of flow to use agents to manage my business. Started with LLM skills that could be used to handle tasks I used to do myself, and since then I've increasingly been having AI use those skills on its own without me invoking them and chain things together into full blown workflows. Some parts I'm still supervising closely, but others that have been working consistently for a while I now don't really watch unless Claude flags something for me to review on my dashboard.

100% agree with this. I think takes like OP's would be much more interesting if they staked out a position in the future. I think it's pretty uncontroversial to say that someone with a great deal of technical expertise is going to be a hugely more effective LLM user today.

The question that really matters is whether that will continue to be the case. My guess is that technical expertise matters less over time, and the ability to specify the desired outcome is eventually the only thing that becomes important. But I could be wrong! The direction this all goes is pretty fuzzy in my mind.

I think the problem with this logic is it's based on the capabilities of LLMs today and really fails to address the prospect that they will continue to improve.

I used to be a PM and am technically literate enough but can only very minimally write code. I have been using LLMs to build (or try to, at least) internal tools for my business since GPT-4.

In the early days, I'd get a little ways, then the LLM would start breaking things, and I'd try but fail to get it to fix things. But over successive generations, I was increasingly able to get it unstuck by offering suggestions on where it may have gone wrong. With Opus 4.7, I don't even really have to do that - if something isn't working it's usually sufficient to just tell it what's broken. It can figure out how to fix it without my input. And of course fewer things are broken in the first place.

So I think I'm very well positioned to understand how these things are improving - better able to get the LLM to do what I want than the post OP quoted from /vibecoding (though I am 99% sure that post is actually AI slop), but less so than most of the people posting in this thread. As they've improved, whatever ability I have to guess at the causes of problems based on my experience having seen things go wrong with products I've PMed has become less necessary to getting the right outcome.

I expect that trend to continue - increasingly the LLM won't need the guidance of people with a great deal of technical expertise. I basically no longer have to attempt to diagnose problems in order to get them fixed, though with the caveat that I am building internal tools for which I am the only user, so certainly much simpler in scope than the stuff OP is talking about.

Without guidance, LLMs tend to paint themselves into a corner, because they’re generating code to solve individual prompts, not thinking holistically about an application’s architecture.

The crux of what I'm trying to say here is that I absolutely believe that this line is 100% true today, but I would be deeply cautious about assuming that it will continue to be true given the improvements in LLMs over the past few years.

If we assume that there will be an AI that is perfect in terms of ability to find vulnerabilities, cheap to run and widely available to everyone, then anyone can run it on any piece of software before deploying it. All vulnerabilities get found before they can be exploited.

One of the big challenges with cybersecurity is that attackers only need to find one exploit, while defenders need to stop everything. When you have a large surface area and limited resources, it's much easier to be the side that only has to succeed once. AI eliminates the limited resources problem.

Nope, wrong again - SOTA LLMs get the car wash thing right.

If you're going to have strong opinions on a topic, it would behoove you to keep up with it.

I'm not. I understand the difference and also that through improvements to the core models as well as harnesses, LLMs are able to handle an increasing share of tasks. I also understand that these things will continue to improve until AI can automate entire jobs.

You, on the other hand, are confusing LLMs from the past with current SOTA LLMs, which can tell how many rs are in strawberry.

It seems likely to me that doctors whose job is almost or entirely about making diagnoses and prescribing treatments won't be able to keep up in the long run, where those who are more patient facing will still be around even after AI is better than us at just about everything.

If I were picking a specialty now, I'd go with pediatrics or psychiatry over something like oncology.