You are in danger. Unless you estimate the odds of a breakthrough at <5%, or you already have enough money to retire, or you expect that AI will usher in enough prosperity that your job will be irrelevant, it is straight-up irresponsible to forgo making a contingency plan.
Personally, rather devolving into nihilism, I'd rather try to hedge against suffering that fate. Now is the time to invest and save money. (or yesterday)
Unless you’re investing in guns, ammo, food, and a bunker. We’re talking worse unemployment than depression era Germany. And structurally more significant unemployment because the people losing their jobs were formally very high earners.
I do also think the mid level bad outcome isn’t super likely because of AI is good enough to replace a lot of white collar jobs, I think it could replace almost all of them.
I would like to know if there's some kind of inflection point, like the so-called Laffer curve for taxes, where once an economy has X% unemployment, it effectively collapses. I'd imagine it goes: recession -> depression -> systemic crisis and appears to be somewhere between 30-40% unemployment based on history.
I do. Show me any evidence that it is imminent.
> or you expect that AI will usher in enough prosperity that your job will be irrelevant
Not in my lifetime.
> it is straight-up irresponsible to forgo making a contingency plan.
No, I'm actually measuring the risk, you're acting as if the sky is falling. What's your contingency plan? Buy a subscription to the revolution?
I’ve been working on my contingency plan for a year-and-a-half now. I won’t get into what it is (nothing earth shattering) but if you haven’t been preparing, I think you’re either not paying enough attention or you’re seriously misreading where this is all going.
Yes I said the word that none of these company want to say in their press conference.
Besides that, why aren't we seeing any metrics change on Github? With a supposedly increase of productivity so large a good chunk of the workforce is fired, we would see it somewhere.
It's not the odds of the breakthrough, but the timeline. A factory worker could have correctly seen that one day automation would replace him, and yet worked his entire career in that role.
There have been a ton of predictions about software engineers, radiologists, and some other roles getting replaced in months. Those predictions have clearly been not so great.
At this point the greater risk to my career seems to be the economy tanking, as that seems to be happening and ongoing. Unfortunately, switching careers can't save you from that.
What contingencies can you really make?
Start training a physical trade, maybe.
If this the end of SWE jobs, you better ride the wave. Odds are you're estimate on when AI takes over are off by half a career, anyways.
I'm baffled that so many people think that only developers are going to be hit and that we especially deserve it. If AI gets so good that you don't need people to understand code anymore, I don't know why you'd need a project manager anymore either, or a CFO, or a graphic designer, etc etc. Even the people that seem to think they're irreplaceable because they have some soft power probably aren't. Like, do VC funds really need humans making decisions in that context..?
Anyway, the practical reason why I'm not screaming in terror right now is because I think the hype machine is entirely off the rails and these things can't be trusted with real jobs. And honestly, I'm starting to wonder how much of tech and social media is just being spammed by bots and sock puppets at this point, because otherwise I don't understand why people are so excited about this hypothetical future. Yay, bots are going to do your job for you while a small handful of business owners profit. And I guess you can use moltbot to manage your not-particularly-busy life of unemployment. Well, until you stop being able to afford the frontier models anyway, which is probably going to dash your dream of vibe coding a startup. Maybe there's a handful of winners, until there's not, because nobody can afford to buy services on a wage of zero dollars. And anyone claiming that the abundance will go to everyone needs to get their head checked.
It seems kind of like saying “I’m smarter than all the AIs in this one particular way.” If someone posted that, you would probably jump in to say they’re fooling themselves.
More likely they get fired for no reason, never rehired, and the people left get burned out trying to hold it all together.
If you fail as a "higher up" you're no longer higher up. Then someone else can take your place. To the extent this does not naturally happen is evidence of petty or major corruptions within the system.
The large, overwhelming majority of my team's time is spent on combing through these tickets and making sense of them. Once we know what the ticket is even trying to say, we're usually out with the solution in a few days at most, so implementation isn't the bottleneck, nowhere near.
This scenario has been the same everywhere I've ever worked, at large, old institutions as well as fresh startups.
The day I'll start worrying is when the AI is capable of following the web of people involved to translate what the vaguely phrased ticket that's been backlogged for God knows how long actually means
However as you point out we have no program-accessible source of data on who stakeholders, contributors, managers, etc. are and have to write a lot of that ourselves. For a smaller business perhaps one could write all of that down in an accessible way to improve this but for a large dynamic business it seems very difficult.
I've been doing stuff with recent models (gemini 3, claude 4.5/6, even smaller, open models like GLM5 and Qwen3-coder-next) that was just unthinkable a few months back. Compiler stuff, including implementing optimizations, generating code to target a new, custom processor, etc. I can ask for a significant new optimization feature in our compiler before going to lunch and come back to find it implemented and tested. This is a compiler that targets a custom processor so there is also verilog code involved. We're having the AI make improvements on both the hardware and software sides - this is deep-in-the-weeds complex stuff and AI is starting to handle it with ease. There are getting to be fewer and fewer things in the ticket tracker that AI can't implement.
A few months ago I would've completely agreed with you, but the game is changing very rapidly now.
I don't agree they have solved this problem, at all, or really in any way that's actually usable.
What helps is also getting it to generate a docs of your code so that it has map.
This is actually how humans understand a large code base too. We don’t hold a large code base in memory — we navigate it through docs and sampling bits of code.
If the answer is that the AI cranks out code faster than the team can digest and review it and faster than you can spec out the features, what’s the point? I can see completely shifting your workflow, letting skills atrophy, adopting new dependencies, and paying new vendors if it’s boosting your final output 5 or 10x.
But if it’s a 20% speed up is it worth it?
In my case the AI is actively detrimental unless I hand hold it with every single file it should look into, lest it dive into weird ancient parts of the codebase that bear no relevance to the task at hand. Letting the latest and "greatest" agents loose is just a recipe for frustration and disaster despite lots of smart people trying their hardest to make these infernal tools be of any use at all. The best I've gotten out of it was some light Vue refactoring, but even then despite AGENTS.md, RULES.md and all the other voodoo people say you should do it's a crapshoot.
It's hard to predict how quickly it will be solved and by whom first, but this appears to be a software engineering problem solvable through effort and resources and time, not a fundamental physical law that must be circumvented like a physical sciences problem. Betting it won't be solved enough to have an impact on the work of today relatively quickly is betting against substantial resources and investment.
Wonder what that means for meatspace.
Edit: Would also disagree this isn’t a physics problem. Pretty sure power required scales according to problem complexity. At a certain level of problem complexity we’re pretty much required to put enough carbon in the atmosphere to cook everyone to a crisp.
Edit 2: illustrative example, an Epic in Jira: “Design fusion reactor”
Plenty of things get substantial resources and investment and go nowhere.
Of course I could be totally wrong and it's solved in the next couple years, it's almost impossible to make these predictions either way. But I get the feeling people are underestimating what it takes to be truly intelligent, especially when efficiency is important.
Well that is easily disproved by the fact that people have children with higher IQ's than their own.
Not to say we can't create machines that far surpass our abilities on a single or small set of axis.
SOTA models are superhuman in a narrow sense, in that they have solid background knowledge of pretty much any subject they've been trained on. That's great. But no, it doesn't turn your AI datacenter into "a country of geniuses".
First of all, it's not necessary for one person to build that super-intelligence all by themselves, or to understand it fully. It can be developed by a team, each of whom understands only a small part of the whole.
Secondly, it doesn't necessarily even require anybody to understand it. The way AI models are built today is by pressing "go" on a giant optimizer. We understand the inputs (data) and the optimizer machine (very expensive linear algebra) and the connective structure of the solution (transformer) but nobody fully understands the loss-minimizing solution that emerges from this process. We study these solutions empirically and are surprised by how they succeed and fail.
We may find we can keep improving the optimization machine, and tweaking the architecture, and eventually hit something with the capacity to grow beyond our own intelligence, and it's not a requirement that anyone understands how the resulting model works.
We also have many instances in nature and history of processes that follow this pattern, where one might expect to find a similar "law". Mammals can give birth to children that grow bigger than their parents. We can make metals puter than the crucible we melted them in. We can make machines more precise than the machines that made those parts. Evolution itself created human intelligence from the repeated application of very simple rules.
Yes, it seems likely to me.
It seems like the ultimate in hubris to assume we are capable of creating something we are not capable of ourselves.
So to create something that exceeds our capabilities is not a matter of hubris (as if physical laws cared about hubris anyway), it's an unambiguously ordinary occurrence.
This is no different than onboarding a new member of the team, and I think openAI was working on that "frontier"
>We started by looking at how enterprises already scale people. They create onboarding processes. They teach institutional knowledge and internal language. They allow learning through experience and improve performance through feedback. They grant access to the right systems and set boundaries. AI coworkers need the same things.
And tribal knowledge will not be a moat once execs realize that all they need to do is prioritize documentation instead of "code velocity" as a metric (sure any metric gets goodhearted, but LLMs are great at sifting through garbage to find the high perplexity tokens).
>But context limitation is fundamental to the technology in its current form
This may not be the case, large enough context-windows plus external scratchpads would mostly obviate the need for true in context learning. The main issue today is that "agent harnesses" suck. The fact that claude code is considered good is more an indication of how bad everything else is. Tool traces read like a drunken newb brute-forcing his way through tasks. LLMs can mostly "one-shot" individual functions, but orchestrating everything is the blocker. (Yes there's progress in metr or whatever but I don't trust any of that, else we'd actually see the results in real-world open source projects).
LLMs don't really know how to interact with subagents. They're generally sort of myopic even with tool calls. They'll spend 20 minutes trying to fix build issues going down a rabbit hole without stepping back to think. I think some sort of self-play might end up solving all of these things, they need to develop a "theory of mind" in the same way that humans do, to understand how to delegate and interact with the subagents they spawn. (Today a failure case is agents often don't realize subagents don't share the same context.)
Some of this is certainly in the base model and pretraining, but it needs to be brought out in the same way RL was needed for tool use.
I look at my ticket tracker and I see basically 100% of it that can be done by AI. Some with assistance because business logic is more complex/not well factored than it should be, but most of the work that is done AI is perfectly capable of doing with a well defined prompt.
Why aren’t there 10x the number of games on steam? Why aren’t people releasing new integrated programming language/OS/dev environments?
Why does our backlog look exactly the same as when I left for posterity leave 4 months ago?
I’m still waiting for the evidence. I still haven’t seen externally verifiable evidence that AI is a net productivity boost for the ability to ship software.
That doesn’t mean that it isn’t. It does mean that it isn’t big enough to be obvious.
I’m very closet watching every external metric I can find. Nothing yet. Just saw the steam metrics for January. Fewer titles than January last year.
That's a sign that you have spurious problems under those tickets or you have a PM problem.
Also, a job is a not a task- if your company has jobs which is a single task then those jobs would definitely be gone.
Live stream validation results as they come in
- What is the validation system and how does it work today?
- What sort of UX do we want? What are the specific deficiencies in the current UX that we're trying to fix?
- What prior art exists on the backend and frontend, and how much of that can/should be reused?
- Are there any scaling or load considerations that need to be accounted for?
I'll probably implement this as 2-3 PRs in a chain touching different parts of the codebase. GPT via Codex will write 80% of the code, and I'll cover the last 20% of polish. Throughout the process I'll prompt it in the right direction when it runs up against questions it can't answer, and check its assumptions about the right way to push this out. I'll make sure that the tests cover what we need them to and that the resultant UX feels good. I'll own the responsibility for covering load considerations and be on the line if anything falls over.
Does it look like software engineering from 3 years ago? Absolutely not. But it's software engineering all the same even if I'm not writing most of the code anymore.
This cyborg process is exactly how we're using AI in our organisation as well. The human in the loop understands the full context of what the feature is and what we're trying to achieve.
For the UX, have it explore your existing repos and copy prior art from there and industry standards to come up with something workable.
Web scale issues can be inferred by the rest of the codebase. If your terraform repo has one RDS server, vs a fleet of them, multi-region, then the AI, just as well as a human, can figure out if it needs Google Spanner level engineering or not. (probably not)
Bigger picture though, what's the process of a human logs an under specified ticket and someone else picks it up and has no clue what to do with it? They're gonna go ask the person who logged the bug for their thoughts and some details beyond "hurr Durr something something validation". If we're at the point where AI is able to make a public blog post shaming the open source developer for not accepting a patch, throwing questions back to you in JIRA about the details of the streaming validation system is well within its capabilities, given the right set of tools.
Ticket trackers are perfect for this. Just start with asking AI to take this unclear, ambiguous ticket and come up with a real plan for how to accomplish it. Review the plan, update your ticket system with the plan, have coworkers review it if you want.
Then when ready, kick off a session for that first phase, first PR, or the whole thing if you want.
Funny enough, all the questions that you posed are things that come up right away that the agent asks itself, and then goes and tries to understand and validate an answer, sometimes with input from the user. But I think this planning mechanism is really critical to being able to have an AI generate an understanding, then have it be validated by a human before beginning implementation.
And by planning I don't necessarily mean plan mode in your agent harness of choice. We use a custom /plan skill in Claude Code that orchestrates all of this using multiple agents, validation loops, and specific prompts to weed out ambiguities by asking clarifying questions using the ask user question tool.
This results in taking really fuzzy requirements and making them clear, and we automate all of this through linear but you could use your ticket tracker of choice.
There will be 3 “software” companies left. And shortly after that society will collapse because of AI can do that it can do any white collar job.
I think it's more nuanced than that. I'd say that - 0% can't be implemented by AI - but a lot of them can be implemented much faster thanks to AI - a lot of them can be implemented slower when using AI (because author has to fix hallucinations, revert changes that caused bugs)
As we learn to use these tools, even in their current state, they will increase productivity by some factor and reduce needs for programmers.
I have seen numerous 25-50% productivity boosts over my career. Not a single one of them reduced the overall need for programmers.
I can’t even think of one that reduced the absolute number of programmers in a specific field.
It's a coding agent that takes a ticket from your tracker, does the work asynchronously, and replies with a pull request. It does progressively understand the codebase. There's a pre-warming step so it's already useful on the first ticket, but it gets better with each one it completes.
The agent itself is done and working well. Right now I'm building out the infrastructure to offer it as a SaaS.
If anyone wants to try it, hit me up. Email is in my profile. Website isn't live yet, but I'm putting together a waitlist.
Um, you do realize that "the memory" is just a text file (or a bunch of interlinked text files) written in plain English. You can write these things out yourself. This is how you use AI effectively, by playing to its strengths and not expecting it to have a crystal ball.