AI agents: Less capability, more reliability, please
I think this piece is underestimating the difficulty involved here though. If only it was as easy as "just pick a single task and make the agent really good at that"!
The problem is that if your UI involves human beings typing or talking to you in a human language, there is an unbounded set of ways things could go wrong. You can't test against every possible variant of what they might say. Humans are bad at clearly expressing things, but even worse is the challenge of ensuring they have a concrete, accurate mental model of what the software can and cannot do.
I can't imagine we're anywhere even close to the kind of perfection required not to need something like this - if it's even possible. Humans use all kinds of review and audit processes precisely because perfection is rarely attainable, and that might be fundamental.
It is almost impossible to produce a useful result, far as I’ve seen, unless one eliminates that mistake from the context window.
There are so many times where I get to a point where the conversation is finally flowing in the way that I want and I would love to "fork" into several directions from that one specific part of the conversation.
Instead I have to rely on a prompt that requests the LLM to compress the entire conversation into a non-prose format that attempts to be as semantically lossless as possible; this sadly never works as in ten did [sic].
If it screws something up it’s highly prone to repeating that mistake
Certainly true, but coaching it past sometimes helps (not always).
- roll back to the point before the mistake.
- add instructions so as to avoid the same path. "Do not try X. We tried X it does not work as it leads to Y.
- add resources that could aid a misunderstanding (api documentation, library code)
- rerun the request (improve/reword with observed details or insights)
I feel like some of the agentic frameworks are already including some of these heuristics, but a helping hand still can work to your benefit
Perhaps the solutions(s) needs to be less focusing on output quality, and more on having a solid process for dealing with errors. Think undo, containers, git, CRDTs
LLMs are supposed to save us from the toils of software engineering, but it looks like we're going to reinvent software engineering to make AI useful.
Problem: Programming languages are too hard.
Solution: AI!
Problem: AI is not reliable, it's hard to specify problems precisely so that it understands what I mean unambiguously.
Solution: Programming languages!
When smartphones first popped up, browsing the web on them was a pain. Now pretty much the whole web has phone versions that make it easier*.
*I recognize the folly of stating this on HN.
There's apps that open links in their embedded browser where ads aren't blocked. So I need to copy the link and open them in my real browser.
Unless these tools can be run locally independent of a service provider, we're just trading one boss for another.
Unless these tools can be run locally independent of a service provider, we're just trading one boss for another.
Many of them already can be. Many more existing models will become local options if/when RAM prices decline.
But this won't necessarily prevent enshittification, as there's always a possibility of a new model being tasked with pushing adverts or propaganda. And perhaps existing models already have been — certainly some people talk as if it's so.
The same fate may come to AI, and that worries me. It won't matter whether you're using OpenAI models, Anthropic models, or locally run models, any more than it matters whether you use Firefox, Chrome or raw cURL - if the business gets to differentiate further between users and AI agents working as users, and especially if they get legal backing to doing that, you can kiss all the benefits of LLMs goodbye, they won't be yours as end-user, they'll all accrue to capitalists, who in turn will lend slivers of it to you, for a price of a subscription.
Any browser you ever used could issue the same HTTP calls (up to standards of a given time, ofc.) - but it helps you with nothing if the endpoint only works when you've signed a contract to access the private API.
Oh, you mean like everyone who shows up to the Cloudflare submissions pointing out how they've been blocklisted from about 50% of the Internet, without recourse, due to the audacity to not run Chrome? In that circumstance, it's actually worse(?) because to the best of my knowledge I cannot subscribe to Cloudflare Verified to not get the :fu: I just have to hope the Eye of Sauron doesn't find me
That reminds me, it's probably time for my semi-annual Google Takeout
Unless these tools can be run locally independent of a service provider, we're just trading one boss for another.
Not only that, we have to be careful about all the integrations being built around it. Thankfully the MCP standard is becoming mainstream (used by Anthropic, OpenAI and next could be Google) and it's an open standard, even if started by Anthropic so we won't have e.g. Anthropic specific integrations.
Once performance of a local setup is on par with online ones or good enough, that'll be game over for them.
So true, it's not going to be "I use PolishDude20GPTBook because my family and friends are on it". It's going to be, "I use PolishDude20GPTBook because they have contracts with Gazeta.pl, Onet, TVN24, OLX and Allegro, so I can use it to get local news and find best-priced products in a convenient way, whereas I can't use TeMPOraLxity for any of that".
Contracts over APIs, again.
As long as the "think of my copyright / AI slop oneoneone" crowd wins. It must not.
They don't want normal people to do this
That's the root of the problem. That's precisely why computers are not the "bicycles for the minds" they were imagined to be.
It's not a conspiracy theory, either. Most of the tech industry makes money inserting themselves between you and your problem and trying to make sure you're stuck with them.
I can’t think of any technological reasons why every digital system can’t have an API (barring security concerns, as those would need to be case by case)
So instead, we put 100s of billions of dollars into statistical models hoping they could do it for us.
It’s kind of backwards.
If I run a flight aggregator that has a majority of flight bookings, I can start charging 'rents' by allowing featured/sponsored listings to be promoted above the 'best' result, leading to a prisoner's dilemma where airlines should pay up to their margins to keep market share.
If an AI company becomes the default application human interface, they can do the same thing. Pay OpenAI tribute or be ended as a going concern.
What I’m saying is that if there was a standard protocol for making travel plans over the internet, we wouldn’t need an AI agent to book a trip.
We could just create great user experiences that expose those APIs like we do for pretty much everything on the web.
The problem is that if your UI involves human beings typing or talking to you in a human language, there is an unbounded set of ways things could go wrong. You can't test against every possible variant of what they might say.
It's almost like we really might benefit from using the advances in AI for stuff like speech recognition to build concrete interfaces with specific predefined vocabularies and a local-first UX. But stuff like that undermines a cloud-based service and a constantly changing interface and the opportunities for general spying and manufacturing "engagement" while people struggle to use the stuff you've made. And of course, producing actual specifications means that you would have to own bugs. Besides eliminating employees, much interest in AI is all about completely eliminating responsibility. As a user of ML-based monitoring products and such for years.. "intelligence" usually implies no real specifications, and no specifications implies no bugs, and no bugs implies rent-seeking behaviour without the burden of any actual responsibilities.
It's frustrating to see how often even technologists buy the story that "users don't want/need concrete specifications" or that "users aren't smart enough to deal with concrete interfaces". It's a trick.
concrete interfaces with specific predefined vocabularies and a local-first UX
An app? We don’t even need to put AI in it, turns out you can book flights without one.
Similarly, software eating the world was actually pretty much fine, but SaaS is/was a bit of a trap. And anyone who thought SaaS was bad should be terrified about the moats and platform lock-in that billion dollar models might mean, the enshittification that inevitably follows market dominance, etc.
Honestly we kinda need a new Stallman for the brave new world, someone who is relentlessly beating the drum on this stuff even if they come across as anticorporate and extreme. An extremist might get traction, but a call to preserve things as they are probably cannot / should not.
More to the point though.. at the beginning at least, Stallman was a respected hacker and not just some random person pushing politics on a community he was barely involved with. It's gotta be that way I think, anyone who's not a respected AI/ML insider won't get far
And.. if you advocate for this I think you risk losing credibility
It's a shame if new interface = credible by default. Look at all the car manufacturers (well some, probably not enough) finally after many years conceding the change to touch interfaces "because new" was a terrible idea, when the right old tool for the job was simply better...and obvious to end-users very quickly.
for general spying and manufacturing "engagement"
Oh, there's one tiny feature that management is really really interested in, make the AI gently upsell the user on a higher tier of subscription if an opportunity presents itself.
And like, as a socially anxious millennial, no I don't particularly like phone calls. However I also recognize that setting my discomfort aside, a direct connection to a human being who can help reason out a problem I'm having is not something easily replaced with a chatbot or an AI assistant. It just isn't. Perfect example: called a place to make a reservation for myself, my wife and girlfriend (poly long story) and found the place didn't usually do reservations on the day in question, but the person did ask when we'd be there. As I was talking to a person, I could provide that information immediately, and say "if you don't take reservations don't worry, that's fine," but it was an off-busy hour so we got one anyway. How does an AI navigate that conversation more efficiently than me?
As a techie person I basically spend the entire day interacting with various software to perform various tasks, work related and otherwise. I cannot overstate: NONE of these interactions, not a single one, is improved one iota by turning it into a conversation, verbal or text-based, with my or someone else's computer. By definition it makes basic tasks take longer, every time, without fail.
Like, really what you're wanting is legitimate information not bound to the whims of advertisers and marketers (and again, to be clear, don't we fucking all) but I don't think an LLM is going to do that for you. If it does it now, and that's a load-bearing if, I have a strong feeling that's because this tech, like all tech, is in it's infancy stage. It hasn't yet gotten enough attention from corporations and their slimy marketing divisions, but that's a temporary state of affairs and has been for every past tech too. Like, OpenAI just closed another funding round and it's valuation is now THREE HUNDRED BILLION. Do you REALLY think they and by extension/as a result, their competitors, are going to be thinking about editorial independence when existing established information institutions already can't?
We need to go back to a more innocent time when we could ask a select group of friends and their trusted chain of friends for recommendations. Not what social media is today.
I dislike driving through Texas, and so, most road trips involve McDonalds - the only time I eat the junk.
My car's inbuilt nav is 13 years out of date, so it knows major throughways but not, for instance, that the road I live on has its own interface to the "highway", and so on, up to restaurants. Phones are unreliable in a lot of the US, and at one point I had a spare phone with all of its storage dedicated to offline Google maps just so I wouldn't get stuck in the Rockies somewhere.
Microsoft used to sell trip planning software and those were the good old days.
My parents were way older than boomers, and hers were boomers, so maybe that's it?
Or, for that matter, solutions you can trust. Remember the pitch for Amazon Dash buttons, where you press it and it maybe-reorders a product for delivery, instantly and sight-unseen? What if the price changed? What if it's not exactly the same product anymore? Wait, did someone else already press it? Maybe I can get a better deal? etc.
Actually, that spurs a random thought: Perhaps some of these smart-speaker ordering pitches land differently if someone is in a socioeconomic class where they're already accustomed to such tasks being done competently by human office-assistants, nannies, etc. Their default expectation might be higher, and they won't need to invest time pinching pennies like the rest of us.
The only problem with most of the flights I book now is that they're with low cost airlines and packed with dark patterns designed to push upgrades.
Would an AI salesman be any better though? At least the website can't actively try to pursuade me to upgrade.
An actually useful agent is something that is totally doable with technologies even from a decade ago, which you by necessity need to host yourself, with a sizeable amount of DIY and duct tape, since it won’t be allowed to exist as a hosted product. The purveyor of goods and services cannot bargain with it so it puts useless junk into your shopping cart on impulse. You cannot really upsell it, all the ad impressions are lost on it, and you cannot phish it with ad buttons that look like the UI of your site — it goes in with the sole purpose to make your bookings/arrangements, it’s a quick in-and-out. It, by its very definition and design, is very adversarial to how most companies with Internet presences run things.
It doesn't have to be this way. Even before the pandemic I remember some companies simply gave me access to an internal app to choose flights where the only flights shown are these of the right date, right airport, and right price.
I think this highlights how we still haven’t cracked intelligence. Many of these issues come from the model’s very limited ability to adapt on the fly.
If you think about it every little action we take is a micro learning opportunity. A small-scale scientific process of trying something and seeing the result. Current AI models can’t really do that.
Yet too many AI projects consistently underestimate this, chasing flashy agent demos promising groundbreaking capabilities—until inevitable failures undermine their credibility.
This is the problem with the 'MCP for Foo' posts that recently.
Adding a capability to your agent that the agent can't use just gives us exactly that:
inevitable failures undermine their credibility
It should be relatively easy for everyone to agree that giving agents an unlimited set of arbitrary capabilities will just make them terrible at everything; and that promising that giving them these capabilities will make them better is:
A) false
B) undermining the credibility of agentic systems
C) undermining the credibility of the people making these promises
...I get it, it is hard to write good agent systems, but surely, a bunch of half-baked, function-calling wrappers that don't really work... like, it's not a good look right?
It's just vibe coding for agents.
I think it's quite reasonable to be say, if you're building a system, now, then:
The key to navigating this tension is focus—choosing a small number of tasks to execute exceptionally well and relentlessly iterating upon them.
^ This seems like exceptionally good advice. If you can't make something that's actually good by iterating on it until it is good and it does work, then you're going to end up being a devin (ie. over promised, over hyped failure).
Love if it could help with that but I haven't figured it out with Google Flights yet. My dream is to tell an AI agent the above and let it figure out the best deal.
Yeah, the "book a flight" agent thing is a running joke now
I literally sat in a meeting with one of our board members who used this exact example of how "AI can do everything now!" and it was REALLY hard not to laugh.
If i were rich enough to have some bot fly me somewhere, I'd have a real-life minion do it for me.
There’s no way I could do what some of these “vibe coders” are doing where they allow AI to write code for them that they don’t even understand.
Basically, what's worse? "Vibes" code that no one understands or a cascade of 20 spreadsheets that no one understands? At least with the "vibes" code you can stick it in git and have some semblance of sane revision control and change tracking.
I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain.
I think this is a great use case for AI, but the analyst still needs to understand what the code that is output does. There are a lot of ways to transform data that result in inaccurate or misleading results.
Vibe coders focus on writing tests
From the boasting I've seen, Vibe coders are also using AI to slop out their tests as well.
I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain.
That sort of makes sense, but then again... if you run some analysis code and it spits out a few plots, how do you know what you're looking at is correct if you have no idea what the code is doing?
Does it reaffirm the biases of the one who signs my paychecks? If so, then the code is correct.
I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain
Considering the hallucinations we've all seen I don't know how they can be comfortable using AI generated data analysis to drive the future direction of the business.
what's worse? "Vibes" code that no one understands or a cascade of 20 spreadsheets that no one understands? At least with the "vibes" code you can stick it in git and have some semblance of sane revision control and change tracking.
You can for spreadsheets too.
think there's a lot of code that gets written that's either disposable or effectively "write only" in that no one is expected to maintain it. I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain.Basically, what's worse? "Vibes" code that no one understands or a cascade of 20 spreadsheets that no one understands?
Correction: it's a "cascade of 20 spreadsheets" that one person understood/understands.
Write only code still needs to work, and someone at some point needs to understand it well enough to know that it works.
A lot of these vibe coders just have a much lower bar for reliability than you.
There's always been a group of beginners that throws stuff together without fully understanding what it does. In the past, this would be copy n' paste from Stackoverflow. Now, that process is simply more automated.
We just need to be better about making it clear which code is that way and which is not.
Check the official website, compare pricing with aggregator, check other dates, check people's availability on cheap dates. Sometimes I only do the first step if the official price is reasonable (I travel 1-2x a month, so I have expectation "how much it should cost").
Don't get me started if I also consider which credit card to use for the points rewards.
But if you wanna find the cheapest way to get to A, compare different retailers, check multiple peoples availability, calculate effects of credit cards etc. It takes time. Aren't those things that could be automated with an agent that can find the cheapest flights, propose dates for it, check availability etc with multiple people via a messing app, calculate which credit card to use, etc?
When I'm picking out a flight I'm looking at, among other things:
* Is the itinerary aggravatingly early or late
* Is the layover aggravatingly short or long
* Is the layover in an airport that sucks
* Is the flight on a carrier that sucks
* What does it cost
If you asked me to encode ahead of time the relative value of each of these dimensions I'd never be able to do it. Heck, the relative value to me isn't even constant over time. But show me five options and I can easily select between them. A clear case where search is more convenient than some agent doing it for me.
It's the same problem with Alexa. I don't trust it to blindly reorder me basic stuff when I have to shift through so many bad product listing on the Amazon marketplace.
An LLM could organise flights with a lower error rate, however, when it goes wrong what is the recourse? I imagine it's anger and a self-promise never to use AI for this again.
*If you're saying that the AI just supplies suggestions then maybe it's useful. Though wouldn't people still be double checking everything anyway? Not sure how much effort this actually saves?
With a really excellent human assistant who deeply understood my brain (at least the travel related parts of it), it was kind of nice. But even then there were times when I thought it would be easier and better to just do it myself. Maybe it's a failure of imagination, but I find it very hard to see the path from today's technology to an AI agent that I would trust enough to hand it off, and that would save enough time and hassle that I wouldn't prefer to just do it myself.
In one chapter he describes his frustration with GPS based navigation apps. I thought it was similar to what you describe.
> If I am commuting home, I may prefer a slower route that avoids traffic jams. (Humans, unlike GPS devices, would rather keep moving slowly than get stuck in stop-start traffic.) GPS devices also have no notion of trade-offs, in particular relating to optimising ‘average, expected journey time’ and minimising ‘variance’ – the difference between the best and the worst journey time for a given route.
For instance, whenever I drive to the airport, I often ignore my GPS. This is because what I need when I’m catching a flight is not the fastest average journey, but the one with the lowest variance in journey time – the one with the ‘least-bad worst-case scenario’. The satnav always recommends that I travel there by motorway, whereas I mostly use the back roads.
Booking a flight is actually task I cannot outsource to a human assistant, let alone AI.
Because there is no "correct" flight. Your preference changes as you discover information about what's available at a given time and price.
The helpful AI assistant would present you with options, you'd choose what you prefer, it would refine the options, and so on, until you make your final selection. There would be no communication lag as there would be with a human assistant. That sounds very doable to me.
For example, this person[0] could have simply booked a United flight from the United site for 15k points. Instead the person batch emailed Turkish Airlines booking offices, found the Thai office that was willing to make that booking but required bank transfers in Thai baht to pay taxes, made two more phone calls to Turkish Airlines to pay taxes with a credit card, and in the end only spent 7.5k points for the same trip on United.
This may be an extreme example, but it shows the amount of familiarity with the points system, the customer service phone tree and the actual rules to get cheap flights.
If AI can do all of that, it'd be useful. Otherwise I'll stick to manual booking.
[0] https://frequentmiler.com/yes-you-can-still-book-united-flig...
Similar to how airlines have tried to stop people exploiting "hidden cities."
This sounds interesting?
Just don't book a round trip, don't check a bag, don't do it too often. Also you're gambling that they don't cancel your flight and book you on a new one to the city you don't actually want to go to (that no longer connects via the hidden city). You can get half price tickets sometimes with this trick.
https://www.bitsaboutmoney.com/archive/seeing-like-a-bank/
As a sophisticated user of the banking system, a useful skill to have is understanding whether the ultimate solution to an issue facing you is probably available to Tier Two or probably only available to a professional earning six figures a year. You can then route your queries to the bank to get in front of the appropriate person with the minimal amount of effort expended on making this happen.You might think bank would hate this, and aggressively direct people who discover side channels to Use The 1-800 Number That Is What It Is For. For better or worse, the side channels are not an accident. They are extremely intentionally designed. Accessing them often requires performance of being a professional-managerial class member or otherwise knowing some financial industry shibboleths. This is not accidental; that greatly cuts down on “misuse” of the side channels by that guy.
All of reality is probabilistic. Expecting that to map deterministically to solving open ended complex problems is absurd. It's vectors all the way down.
Expecting that to map deterministically to solving open ended complex problems is absurd.
TCP creates an abstraction layer with more reliability than what it's built on. If you can detect failure, you can create a retry loop, assuming you can understand the rules of the environment you're operating in.
If you can detect failure, you can create a retry loop, assuming you can understand the rules of the environment you're operating in
Indeed, this is what makes autonomous agentic tool using systems robust as well. Those retry loops become ad-hoc where needed, and the agent can self correct based on error responses, compared to a defined workflow that would get stuck in said loop if it couldn't figure things out, or just error out the whole process.
Stability is the bedrock of the evolution of stable systems.
So we also thought with AI in general, and spent decades toiling on rules based systems. Until interpretability was thrown out the window and we just started letting deep learning algorithms run wild with endless compute, and looked at the actual results. This will be very similar.
”It works when it works” is fine when stakes are low and human is in the loop, like artwork for a blog post. And so in a way, I agree with you. AI doesn’t belong in intermediate computer-to-computer interactions, unless the stakes are low. What scares me is that the AI optimists are desperately looking to apply LLMs to domains and tasks where the cost of mistakes are high.
Flights are a good example but I often cite Uber as a good one too. Nobody wants to tell their assistant to book them an Uber - the UX/UI is so streamlined and easy, it's almost always easy enough to just do it yourself (or if you are too important for that, you probably have a private driver already). Basically anything you can do with an iPhone and the top 20 apps is in this category. You are literally competing against hundreds of engineers/product designers who had no other goal than to build the best possible experience for accomplishing X. Even if LLMs would have been helpful a priori - they aren't after every edge case has already been enumerated and planned for.
You are literally competing against hundreds of engineers/product designers who had no other goal than to build the best possible experience for accomplishing X.
I think part of what's been happening here is that the hubris of the AI startups is really showing through.
People working on these startups are by definition much more likely than average to have bought the AI hype. And what's the AI hype? That AI will replace humans at somewhere between "a lot" and "all" tasks.
Given that we're filtering for people who believe that, it's unsurprising that they consciously or unconsciously devalue all the human effort that went into the designs of the apps they're looking to replace and think that an LLM could do better.
I think part of what's been happening here is that the hubris of the AI startups is really showing through.
I think it its somewhat reductive to assign this "hubris" to "AI startups". I would posit that this hubris is more akin to the superiority we feel as human beings.
I have heard people say several times that they "treat AI like a Jr. employee", I think that within the context of a project AI should be treated based on the level if contribution. If AI is the expert, I am not going to approach it as if I am an SME that knows exactly what to ask. I am going to try and focus on the thing. know best, and ask questions around that to discover and learn the best approach. Obviously there is nuance here that is outside the scope of this discussion, but these two fundamentally different approaches have yield materially different outcomes in my experience.
Absolutely not. When giving tasks to an AI, we supply them with context, examples of what to do, examples of what not to do, and we clarify their role and job. We stick with them as they work and direct them accordingly when something goes wrong.
I've no idea what would happen if we treated a junior developer like that.
The problem I find in many cases is that people are restrained by their imagination of what's possible, so they target existing workflows for AI.
I concur and would like to add that they are also restrained by the limitations of existing "systems" and our implicit and explicit expectations of said system. I am currently attempting to mitigate the harm done by this restriction by focusing on and starting with a first principal analysis of the problem being solved before starting the work, for example; lets take a well established and well documented system like the SSA.
When attempting to develop, refactor, extend etc... such a system; what is the proper thought process. As I see it, there are two paths:
Path 1:
a) Breakdown the existing workflows
b) Identify key performance indicators (KPIs) that align with your business goals
c) Collect and analyze data related to those KPIs using BPM tools
d) Find the most expensive worst performing workflows
e) Automate them E2E w/ interface contracts on either side
Path 2:
a) Analyze system to understand goal in terms of 1st principals, e.g: What is the mission of the SSA? To move money based on conditional logic.
b) What systems / data structures are closest to this function and does the legacy system reflect this at its core e.g.: SSA should just be a ledger IMO
c) If Yes, go to "Path 1" and if No go to "D"
d) Identify the core function of the system, the critical path (core workflow) and all required parties
e) Make MVP which only does the bare min
Note: Obviously this is a gross oversimplification of the project management process and there are usually external factors that weigh in and decide which path is possible for a given initiative, my goal here was just to highlight a specific deviation from my normal process that has yielded benefits so far in my own personal experience.
It's like saying rm -rf / should have more safeguards built in. It feels unfair to call out the AI based tools for this.
* "unreliable" may not be the right word. For all we know, the agent performed admirably given whatever the user's prompt may have been. Just goes to show that even in a relatively constricted domain of programming, where a lot (but far from all) outcomes are binary, the room for misinterpretation and error is still quite vast.
Any system capable of automating a complex task will by need be more complex than the task at hand. This complexity doesn't evaporate when you through statistical fuzzers at it.
For example, if a user with appropriate privileges mistakenly runs ‘rm -rf / tmp/junk’, that may remove all files on the entire system. Since there are so few legitimate uses for such a command, GNU rm normally declines to operate on any directory that resolves to /. If you really want to try to remove all the files on your system, you can use the --no-preserve-root option, but the default behavior, specified by the --preserve-root option, is safer for most purposes.
https://www.gnu.org/software/coreutils/manual/html_node/Trea...
I hypothesize that a $(git fetch --mirror) would pull down the "orphaned" revision, too, but don't currently have the mental energy to prove it
And I'm so sick of everything trying for 100% automation and failing. There's a place for the human in the loop, in quickly identifying bugs the AI doesn't have the context for, or large-scale vision, or security or product-focused mindset, etc.
It's going to be AI and humans collaborating. The solutions that figure that out the best are going to win IMO. AI won't be doing everything and humans won't be doing it all either. The tools with the best human-AI collaboration are where it's at.
For python changes I use ast and astor packages, for JS I use esprima/escodegen/estraverse, and for CSS I use postcss. The process is the same for each one: I give the original input souce file, and I instruct the LLM to parse the file into AST form and then write code that modifies that AST.
I blogged about it here if you want more details! https://codeplusequalsai.com/static/blog/prompting_llms_to_m...
I still think like prompting is still the wrong interface for programming systems. Even though they're restricted, configurations forms, visual programming with nodes, and small scripts attached to objects on a platform is way more reliable and useful.
The best companies can get up to 90% accuracy. Most are closer to 80%.
But it's important to remember, we're expecting perfection here. But think about this: Have you ever asked someone to book a flight for you? How did it go?
At least in my experience, there's usually a few back and forth emails, and then something is always not quite right or as good as if you did it yourself, but you're ok with that because it saved you time. The one thing that makes it better is if the same person does it for you a couple of times and learned your specific habits and what you care about.
I think the biggest problem in AI accuracy is expecting the AI to be better than a human.
I think the biggest problem in AI accuracy is expecting the AI to be better than a human.
If it's not better across at least one of {more accurate, faster, cheaper} then there is no business. You have to be offering one of the above.
And that applies both to humans and to existing tech solutions: an LLM solution must beat both in some dimension. Current flight booking interfaces are actually better than a human at all three: they're more accurate, they're free, and they're faster than trying to do the back and forth, which means the bar to clear for an agent is extremely high.
Current flight booking interfaces are actually better than a human at all three
Only when you know exactly where to go. If you need to get to customers in 3 cities where order doesn't matter (ie the traveling salemen problem, though you are allowed to hit any city more than once) current solutions are not great. If you want to go on vacation but don't care much about where (almost every place with an airport would be an acceptable vacation, though some are better than others)
Does it terrify anyone else that there is an entire cohort of new engineers who are getting into programming because of AI, but missing these absolute basic bare necessities?> Terrify? No, it's reassuring that I might still have a place in the world.
[0] https://www.reddit.com/r/cursor/comments/1inoryp/comment/mdo...
Why would you ask the community a question like "how to source control" when you've been working with (presumably) a programming genius LLM that could provide the most personally tailored path for baby's first git experience? Even if you don't know that "git" is a thing, you could ask questions as if you were a golden retriever and the model would still inevitably recommend git in the first turn of conversation.
Is it really the case that a person who has the ability to use a compiler, IDE, LLM, web browser, reddit, etc., somehow simultaneously lacks the ability to frame basic-ass questions about the very mission they set out on? If stuff like this is not manufactured, then we should all walk away feeling pretty fantastic about our future job prospects.
People think "this is hard, I'll re-invent it in an easier way" and end up with a half-assed version of the tooling we've honed over the decades.
People think "this is hard, I'll re-invent it in an easier way" and end up with a half-assed version of the tooling we've honed over the decades.
This is a win in the long run because the occassional and successful thought people labor over sometimes is a better way.
When the goal is "re-invent programming to make it easier" all you get is a hodgepodge of half-ass solutions like GP said. Enhancing traditional focused workflows seems a lot more interesting to me than "coding assistant".
Hopefully AI tooling will continue to evolve. I don't see how you get around the reliability issues with this iteration of AI (GPT+RLHF+RAG, etc). Transfer learning is still abysmal.
https://www.reddit.com/r/cursor/comments/1inoryp/comment/mdr...
I'm not a dev or engineers at all (just a geek working in Finance)
This fits my experience of teaching very intelligent students how to code; if you're an experienced programmer, you simply cannot fathom the kinds of assumptions beginners will make due to gaps in yet-to-be foundational knowledge. I remember having to tell students to mindful when searching Stack Overflow for help, because of how something as simple as an error from Requests (e.g. while doing web scraping) could lead them down a rabbit hole of "solutions" such as completely uninstalling their Python for a different/older version of Python.
Plug: We just posted a demo of our agent doing sophisticated reasoning over a huge dataset ((JFK assassination files -- 80,000 PDF pages): https://x.com/peterjliu/status/1906711224261464320
Even on small amounts of files, I think there's quite a palpable difference in reliability/accuracy vs the big AI players.
The most important thing is doing rigorous evals that are representative of what your users do in your product. Often this is not the same as academic benchmarks.
OMFG thank you for saying this. As a core contributor to RA.Aid, optimizing it for SWE-bench seems like it would actively go against perf on real-world tasks. RA.Aid came about in the first place as a pragmatic programming tool (I created it while making another software startup, Fictie.) It works well because it was literally made and tested by making other software, and these days it mostly creates its own code.
Do you have any tips or suggestions on how to do more formalized evals, but on tasks that resemble real world tasks?
And before going to crowd-workers (maybe you can skip them entirely) try LLMs.
I would start by making the examples yourself initially
What I'm doing right now is this:
1) I have X problem to solve using the coding agent.
2) I ask the agent to do X
3) I use my own brain: did the agent do it correctly?
The hardest part about automating this is #3 above --each evaluation is one-off and it would be hard to even formalize the evaluation.
SWE bench, for example uses unit tests for this, and the agent is blind to the unit tests --so the agent has to make a red test (which it has never seen) go green.
Rather than blaming the agent, we should recognize that this behavior is expected. It’s not that AI is uniquely flawed — it's that we're automating a class of human communication problems that already exist.
This is less about broken tools and more about adjusting our expectations. Just like hunters had to learn how to manage gunpowder weapons after using bows, we’re now figuring out how to responsibly wield this new power.
After all, when something works exactly as intended, we already have a word for that: software.
+100 on the footnote:
agents or workflows?
Workflows. Workflows, all the way.
The agents can start using these workflows once they are actually ready to execute stuff with high precision. And, by then we would have figured out how to create effective, accurate and easily diagnozable workflows, so people will stop complaining about "I want to know what's going on inside the black box".
99.9% of real world enterprise AI use cases today are for workflows not agents.
However, "agents" are being pushed because the industry needs a next big thing to keep the investment funding flowing in.
The problem is that even the best reasoning models available today don't have the actual reasoning and planning capability needed to build truly autonomous agents. They might in a year. Or they might not.
And are there any guidelines on how to manage workflows for a project or set of projects? I’m just keeping them in plain text and including them in conversations ad hoc.
It’s easy to blame the user's missing grasp of basic version control, but that misses the deeper point.
Uhh, no, that's pretty much the point. A developer without basic understanding of version control is like a pilot without a basic understanding of landing. A ton of problems with AI (or any other tool, including your own brain) get fixed by iterating on small commits and branching. Throw away the commit or branch if it really goes sideways. I can't fathom working on something for 4 months without realizing a problem or having any way to roll back.
That said, the one argument I could see is if Cursor (or copilot, etc) had built in to suggest "this project isn't in source control, we should probably do that before getting too far ahead of ourselves.", then help the user setup sc, repo, commit, etc. The topic _is_ tricky and I do remember not totally grasping git, branching, etc.
Here's CNBC Business is pushing greed that these aren't AI wrappers but next best thing after fire, bread and axe[0]
Often when using an AI agent, I think to myself that a web search gets me what I need more reliably and just as quick. Maybe AI has to learn to crawl before it learns to walk, but each agent I use is leaving me without confidence that it will ever be useful and I genuinely wonder if they've ever been tested before being published...
Nowhere in that story above is there a customer or factory worker feeding in open-ended inputs. The factory is precise, it takes inputs and produces outputs. The variability is restricted to variability of inputs and the reliability of the factory kit.
Much business software is analogous to the factory. You have human workers who ultimately operate the business. And software is built to automate those tasks precisely.
AI struggles because engineers are trying to build factories through incantation - if they just say the right series of magic spells, the LLM will produce a factory.
And often it can. It’s just a shitty factory that does simple things, often inefficiently with unforeseen edge cases.
At the moment, skilled factory builders (software engineers) are better at holistically understanding the needs of the business and building precise, maintainable, specific factories.
The factory builders will use AI as a tool to help build better factories. Trying to get the AI to build the whole factory soup-to-nuts won’t work.
I tend to think that what this article is asking for isn't achievable, because what people mean by "AI" is precisely "we don't know how it works".
An analogy I've used sometimes when talking with people about AI is the "I know a guy" situation. Someone you know comes and tells you "I know a guy who can do X for you", where "do X" is "write your class paper" or "book a flight" or "describe what a supernova is" or "invest your life savings". In this situation, the more important the task, the more you would probably want to know about this "guy". What are his credentials? Has he done this before? How often has he failed? What were the consequences? Can he be trusted? Etc.
The thing that "a guy" and an AI have in common is that you don't know what they're doing. Where they differ is in your ability to gradually gain knowledge. In real life, "know a guy" situations become transformed into something more specific as you gain information about who the person is and how they do what they do, and especially as you understand more about the system of consequences in which they are embedded (e.g., "if this painter had ruined many people's houses he would have been sued into oblivion, or at least I would have heard about it"). And also real people are unavoidably embedded in the system of physical reality which imposes certain constraints that bound plausibility (e.g., if someone tells you "I know a guy who can paint your entire house in five seconds" you will smell a rat).
Asking for "reliability" means asking for a network of causes and effects that surrounds and supports whatever "guy" or AI you're relying on. At this point I don't see any mechanism to provide that other than social and ultimately legal pressure, and I don't see any strong action being taken in that direction.
That's my experience working with a largeish mature codebase (all on non-prod code) where you can't get far if you can't use various internal libraries correctly. With standalone (or small greenfield) projects, where results can lean more on public info from pre-training and there's not as much project specific info to pull in, you might see different outcomes.
Maybe the tech and surrounding practice will change over time, but in my short experience it's mostly been about trying to just get to 'acceptable' for this kind of task.
I've started taking a very data engineering-centric approach to the problem where you treat an LLM as an API call as you would any other tool in a pipeline, and it's crazy (or maybe not so crazy) what LLM workflows are capable of doing, all with increased reliability. So much so that I've tried to package my thoughts / opinions up into an AI SDK for Apache Airflow[1] (one of the more popular orchestration tools that data engineers use). This feels like the right approach and in our customer base / community, it also maps perfectly to the organizations that have been most successful. The number of times I've seen companies stand up an AI team without really understanding _what problem they want to solve_...
That being said, back in February I was trying out of bunch of AI personal assistant apps/tools. I found, without fail, every single one of them was advertising features their LLMs could theoretically accomplish, but in practice couldn't. Even worse was many of these "assistants" would proactively suggest they could accomplish something but when you sent them out to do it, they'd tell you they couldn't.
* "Would you like me to call that restaurant?"...."Sorry, I don't have support for that yet"
* "Would you like me to create a reminder?"....Created the reminder, but never executed it
* "Do you want me to check their website?"...."Sorry, I don't support that yet"
Of all of the promised features, the only thing I ended up using any of them for was a text message interface to an LLM. Now that Siri has native ChatGPT support, it's not necessary.
When we tried the 'full agent' approach (letting it roam freely through our codebase), we ended up with some impressive demos but constant production incidents. We've since pivoted to more constrained workflows with human checkpoints, and while less flashy, user satisfaction has gone way up.
The Cursor wipeout incident is a perfect example. It's not about blaming users who don't understand git - it's about tools that should know better. When I hand my code to another developer, they understand the implied contract of 'don't delete all my shit without asking.' Why should AI get a pass?
Reliable > clever. It's the difference between a senior engineer who delivers consistently and a junior who occasionally writes brilliant code but breaks the build every other week."
Image compression, eventual consistency, fuzzy search. There are many more examples I'm sure.
Isn't all of these very deterministic? You can predict what's going to be discarded by the compression algorithm. Eventual consistency is only eventual because of the generation of events. Once that stops, you will have a consistent system and the whole thing can be replayed based on the history of events. Even with fuzzy search you can intuit how to get reliable results and ordering without even looking at the algorithms.
An LLMs based agent is the least efficient method for most of the cases they're marketing if for. Sometimes all you need is a rule-based engine. Then you can add bounded fuzziness where it's actually helpful.
As far as I know, llm hallucinations are inherent to them and will never be completely removed. If I book a flight, i want 100,0% reliability, Not 99% ( which we are still far away today).
People got to take llm for what they are, good bullshiter, awesome to translate text or reformulate words but it's not designed to have thought or be an alternate secretary. Merely a secretary tool.
Given the intensifying competition within AI, teams face a difficult balance: move fast and risk breaking things, or prioritize reliability and risk being left behind.
Can we please retire this dichotomy? Part of why teams do this in the first place is because there's this language of "being left behind."
We badly need to retreat to a world in which rigorous engineering is applauded and expected—not treated as a nice to have or "old world thinking."
https://news.ycombinator.com/item?id=43350950
There are mistakes in the Manus demo if you actually look at it. And with so many AI demos, they never want you to look too closely because the thing that was created is fairly mediocre. No one is asking for the tsunami of sludge except for VCs apparently.
It's not like there's a lever in Cursor HQ where one side is "Capability" and one side is "Reliability", and they can make things better just by tipping it back towards the latter.
You can bias designs and efforts in that direction, and get your tool to output reversible steps or bake in sanity checks to blessed actions, but that doesn't change the nature of the problem.
AI agents are not there yet and even cursor has agent mode not selected by default. I have seen cursor agent quite a bit worse that the raw model with human selected context.
Perhaps the question “is this irreversible?” should be delegated to a separate model invocation.
There could be a future in which agentic systems are a tree of model and tool invocations, maybe with a shared scratchpad.
That's why I claim that any sufficiently complicated cognitive architecture contains an ad hoc, informally-specified, bug-ridden, slow implementation of half of Immanuel Kant's work.
I personally struggle to find a new one (AI agent coding assistants already exist, and of course I'm excited about them, especially as they get better). I will not, any time soon, trust unsupervised AI to send emails on my behalf, make travel reservations, or perform other actions that are very costly to fix. AI as a shopping agent just isn't too exciting for me, since I do not believe I actually know what features in a speaker / laptop / car I want until I do my own research by reading what experts and users say.
Transparency? If it worked even unreliably, nobody would care what it does. Problem is stochastic machines aren't engineers, don't reason, are not intelligence.
I find articles attacking Ai but finding excuses in some mouse rather than pointing at the elephant, exhausting.