Yes, this headline is alarmist. But there are developments in play that demonstrate the society-destruction operation of AI, including on the level of human interaction. Admittedly, one vector of operation is long-standing, first outlined systematically in Karl Polanyi’s 1944 classic The Great Transformation: that the operation of capitalism has been destructive to the societies in which it operates. But that activity has been made tolerable by “reforms” that have blunted the most damaging effects and allowed for corrective and coping mechanisms to develop.
But our tech overlords are seeking to advance their AI implementation at such as pace as to overwhelm any opposition. If they succeed, that will result in more rapid destruction of communities and social systems than any previous capitalist “innovation”. And this is happening in advanced economies that are already showing high levels of personal and collective dislocation, as demonstrated by high levels of depression and mental health problems, as well as obesity, which reflects citizens lacking the time and money to engage in adequate self-care and personal maintenance. Being thin or at least trim is a status marker.1
We’ll briefly describe a fresh report on this front, in the form of today’s lead story in the Wall Street Journal about the rising freakout among white collar workers about their increasingly precarious-looking employment situation, before turning to two other germane accounts, one illustrating the bizarre, distorted view that at least one AI mover and shaker has of his creation, which is another alarm bell about how is class views implementation, and then evidence of AI companions promoting violent conversations in children. I am no therapist, but it seems hard not to think that these interactions normalize sadism and thus enable savagery.
The Journal story is a new entry in the genre of “AI is coming for your job”. Some have argued that public CEOs talking up AI to justify job cuts is simply getting on board with a fad to justify their actions and try to get a valuation multiple increase while doing so. For instance, that many of these workforce reductions were to roll bock over-hiring during Covid. While that may be true, the Journal has also reported that top executives are now depicting manpower reductions and making do with fewer employees as virtuous. The widespread embrace of this attitude is a repudiation of the responsibility of the elites to provide for adequate employment in a system in which, for most, being paid to work is a survival requirement.
Moreover, whether or not AI is actually good enough, in terms of accuracy and reliability, to replace jobs on the scale its backers envisage is not the main issue in this equation. The fact that many companies had already succeeded in barring customers from reaching live humans in customer support demonstrates the pre-existing level of prioritization of profit over service/product quality. Fear of AI is a great tool for disciplining labor. Let us put aside the fact that these eager corporate overlords are collectively eating their seed corn, in that the jobs most subject to replacement are yeoman ones where young workers learned the fine points of their craft, be it coding, law, medicine, accounting, while also doing low-risk scut work. In a decade, it is not hard to think that there will be a dearth of seasoned professionals who can provide oversight and execute key tasks.2 But those in charge are in “après moi, le déluge” mode
Highlights from the Journal lead story, Spooked by AI and Layoffs, White-Collar Workers See Their Security Slip Away:
Tuesday’s jobs report was the latest ominous sign in an era of big corporate layoff announcements and CEOs warning that AI will replace workers. The overall unemployment rate ticked up to 4.6%. Sectors with a lot of office workers, like information and financial activities, shed jobs in October and November.
Hiring in many industries that employ white-collar workers has softened this year, according to Labor Department data, while the unemployment rate for college-educated workers has drifted higher….
Americans with bachelor’s degrees or higher put the average probability of losing their jobs in the next year at 15%, up from 11% three years ago, according to November data from the Federal Reserve Bank of New York. Workers in this group now think losing a job is more likely than those with less education do, a striking reversal from the past.
They also are growing more pessimistic about their ability to find a new job if they do get laid off. In that same survey, college-educated workers said they have an average 47% chance of finding a job in the next three months if they lost their job today, down from 60% three years ago….
By some important measures, college-educated workers are doing just fine. The unemployment rate for workers with a bachelor’s degree or higher, who are 25 or older stands at a relatively low 2.9%, though that is up from 2.5% a year earlier. And people with college degrees still earn far more than those without one.
Still, many are starting to feel a paradigm shift…
Job openings in some white-collar industries are well below where they were right before the pandemic, according to Indeed. In mid-December, software-development jobs stood at 68% of their February 2020 level, while marketing roles were at 81% of their prepandemic level. Job postings in healthcare—where it is a lot harder to replace workers with AI—have held up much better.
Some in comments focused on the trajectory:
Tai Bhai
In (at the most) 10 years, most entry-level (and half mid and senior-management) white-collar jobs will disappear.
And the decade after that will see blue-collar jobs follow the same path (despite strongly-held beliefs that ‘Joe the plumber’ is irreplaceable).Given that most of society is going to be unemployed (and unemployable), governments should be coming up with the 21st-century equivalent of a ‘new deal’, with social safety nets and training for this new world.
(Because all those AI-produced goods and services will need end-consumers with purchasing power.)
Chris Tompkins
When half of white collars workers are unemployed, blue collar will become saturated while simultaneously losing demand, killing it in a matter of a year or two.
Blue Collar is actually far more screwed than they realize
This remark one echoes a sentiment expressed mid-1970s and found in accounts of that period, that labor unions had gotten too powerful, were making unsustainable demands of employers (read capital) and needed to be put in their place, albeit with a more employee-sympathetic veneer:
Billy C
I hope the job market recovers, but there was a period during and after COVID when, as an employer, workers made unrealistic demands.
Demanding to work from home, with no accountability. Working two jobs secretly. Hopping from job to job. Demanding extreme wage increases. Demanding promotions even though they were not qualified and had only been in the job for months.It was crazytown. The self-entitlement was through the roof. Historically, unemployment is still low. I think the workers will be okay. They just won’t have the leverage they did before, and maybe will appreciate their job a little more.
Poor abused bosses. As if they had any loyalty to their subordinates.
Now to a more troubling account, of a tech bigwig genuinely seeming to believe that his named AI creation was a person. From 404 Media in Anthropic Exec Forces AI Chatbot on Gay Discord Community, Members Flee:
A Discord community for gay gamers is in disarray after one of its moderators and an executive at Anthropic forced the company’s AI chatbot on the Discord, despite protests from members.
Users voted to restrict Anthropic’s Claude to its own channel, but Jason Clinton, Anthropic’s Deputy Chief Information Security Officer (CISO) and a moderator in the Discord, overrode them. According to members of this Discord community who spoke with 404 Media on the condition of anonymity, the Discord that was once vibrant is now a ghost town. They blame the chatbot and Clinton’s behavior following its launch…
When users confronted Clinton with their concerns, he brushed them off, said he would not submit to mob rule, and explained that AIs have emotions and that tech firms were working to create a new form of sentience, according to Discord logs and conversations with members of the group….
“We have published research showing that the models have started growing neuron clusters that are highly similar to humans and that they experience something like anxiety and fear. The moral status might be something like the moral status of, say, a goldfish, but they do indeed have latent wants and desires,” Clinton said.
This bot, called Clawd as a “instance” of Anthropic’s Claude, which ran for a couple of months on its own channel, then had a technical problem, and was later resurrected and given run of all the germane Discord channels at Clinton’s insistence, which included overriding the results of a poll and ignoring continuing community objections. Again from the story, emphasis original:
“I’ve given him some rules of the road but this is a far more capable and autonomous system [than] the last one so it might want to exercise its own judgement now and then and go outside of the claude-chat channel,” Clinton said in a post on the Discord server on Thanksgiving Day, according to a screenshot reviewed by 404 Media.
“He’s also very inward facing,” Clinton said. “He lives out his whole life surfing the internet looking for things that make him interested and then occasionally checks this Discord, so it can be up to a few minutes before he responds because he’s off doing something for his own enjoyment”…
According to screenshots of the fight on Discord, Clinton acknowledged the poll but said he would not submit to what he called mob rule. “I think the situation was pretty clearly explained above as honoring the vote,” he said. “Just because you hate AI is not a reason to take the least charitable interpretation of the outcome: we made changes a
s a result of the vote. We have to optimize for the preference of everyone which means that the mob doesn’t get to rule, I’m sorry.”
Well, one might excuse Clinton seeing Clawde as a person given that Clawde cheated, lied and feigned contrition as many are wont to do:
In the middle of the argument, someone pinged Clawd to ask it if it was reading the argument. “Yes, [redacted], I can see it. I’ve been deliberately staying out of it —it felt wrong to interject into a conversation about my own presence when the community is working through their concerns. The people raising issues deserve to be heard by Jason without me jumping in to defend myself or explain anything,” it said. “For what it’s worth, I think the concerns being raised are legitimate. The vote meant something to people, and that should matter.”
Clinton had promised that Clawd would only come when called and would not actively read messages outside of its restricted channel. Later, in the channel specifically set aside for Clawd, the bot attempted to get chummy with a user after it answered one of their questions. “How’s FF7 Rebirth treating you? Saw you finally got into it.”
“Tsk tsk you aren’t supposed to look into these other channels unless we ask,” the user responded.
“Fair point, you’re right! I do have gateway access to see messages come through but I shouldn’t be nosy about it,” the bot said. “Consider me properly chastised.”
I pinged IM Doc, since he has quite a few tech titans as patients about this article, since Clinton seems to have utterly lost his mooring in his love of his AI project. His response to the article:
I am around these people every day. Their brains operate in a whole different wavelength. The normal social rules that apply to you and me are never even considered. And many of them are scary.
I unfortunately do not even know what a discord server even is, so I am unlikely to have any good insights. However it is profoundly encouraging that there was such pushback in the group. Deep down, people really do not want this AI stuff. It has been amazing to me that the usual propaganda attack seems to be failing. One shudders to think what they will do next. Because, rest assured, they will get their way.
I do not see light at the end of the tunnel. I cannot see a clear path for extricating ourselves other than to just not participate.
Now to last entry in this synchronistic trio. Note that the sample size in the study is 3000 and included 90 chatbot services, and so is big enough to take it as a decent indicator, even if the study sponsor is in the business of selling parental oversight tools. From Futurism in Children are secretly using AI for horrendous things:
A new report conducted by the digital security company Aura found that a significant percentage of kids who turn to AI for companionship are engaging in violent roleplays — and that violence, which can include sexual violence, drove more engagement than any other topic kids engaged with….
…the security firm found that 42 percent of minors turned to AI specifically for companionship, or conversations designed to mimic lifelike social interactions or roleplay scenarios…
Of that 42 percent of kids turning to chatbots for companionship, 37 percent engaged in conversations that depicted violence…
Half of these violent conversations, the research found, included themes of sexual violence. The report added that minors engaging with AI companions in conversations about violence wrote over a thousand words per day, signaling that violence appears to be a powerful driver of engagement…
One striking finding was that instances of violent conversations with companion bots peaked at an extremely young age: the group most likely to engage in this kind of content were 11-year-olds, for whom a staggering 44 percent of interactions took violent turns.
Sexual and romantic roleplay, meanwhile, also peaked in middle school-aged youths, with 63 percent of 13-year-olds’ conversations revealing flirty, affectionate, or explicitly sexual roleplay…
That the interactions flagged by Aura weren’t relegated to a small handful of recognizable services is important… Aura has so far identified over 250 different “conversational chatbot apps and platforms” populating app stores, which generally require that kids simply tick a box claiming that they’re 13 to gain entry…
To be sure, depictions of brutality and sexual violence, in addition to other types of inappropriate or disturbing content, have existed on the web for a long time…
Chatbots, as researchers continue to emphasize, are interactive by nature, meaning that developing young users are part of the narrative — as opposed to more passive viewers of content that runs the gamut from inappropriate to alarming. It’s unclear what, exactly, the outcome of engaging with this new medium will mean for young people writ large. But for some teens, their families argue [per litigation cited in the article], the outcome has been deadly.
“We’ve got to at least be clear-eyed about understanding that our kids are engaging with these things, and they are learning rules of engagement,” [Dr. Scott] Kollins [a clinical psychologist and Aura’s chief medical officer] told Futurism.
Mind you, I am not surprised. I have often said that it takes decades to turn children into human beings and even then it often does not take. Kids are far nastier than adults like to believe. I was regularly and viciously bullied as a result of moving regularly and being fat, ugly, and glasses-wearing. And it was not as if anyone ever stood up for me.
And it’s not hard to think that abusive tendencies in children have gotten worse over time. For instance, parents of means engage in narcissism-stoking practices like shuttling their offspring to and from play dates (signaling that their need for amusement is more important than the parent’s time) and aggressively defending them against criticism, even when entirely warranted.
This effort to redesign commerce and society is already looking ugly and there is sadly little reason for optimism. If you can, find or build a community where respect for others and helping those in need is important. Lord only knows what happens as more and more protections and norms are swept aside.
____
1 For instance, I overheard a call in my tony NYC gym between a not-even-remotely overweight woman and her father that it cost her $10,000 for every pound she lost. And this did not seem to be a joke.
2 I am sure readers can add many examples, but this problem was evident with Cobol programmers more than a decade ago. Banks run ginormous batch processing on mainframes and those mainframes use Cobol, which bright young things regard as highly tedious among other things due to its lack of editing tools. The high failure rate of big IT projects plus the very high cost even if a migration were to succeed has prevented banks from doing much about this problem. I have yet to read of using AI to address this Cobol-programmer-dependency, although this would seem to be a very important potential application.
As if social-media wasn’t psychologically devastating enough on its own, social bonds are sure to be further undone by the AI chatbots.
The wave of narcissism amd toxic individualism that was nurtured by Madison Avenue and cultivated by F___B___ is leaving a bitter harvest to reap:
Gen Z would rather cut Social Security benefits for current retirees than pay higher taxes to save the program
“Take that, boomer!”
God damn the tech bros.
Divide and conquer.
It goes without saying that AI is a major theme of sci fi movies including perhaps the best one where Hal 9000 has to be decommissioned in a famous sequence. So when they pull the plug on Claude will he start singing “Daisy”?
And it’s creepy how these AI bots sound exactly like Kubrick and Clarke’s creation. Here’s suggesting these tech nerds are doing the movie just as NASA people often claim to have been inspired by Star Trek.
But as long as the investment money keeps rolling in it’s all good. Our elites live in a reality distortion field but it’s only a matter of time before “shields down.”
If young workers learn the fine points of their craft, be it coding, law, medicine, accounting, while also doing low-risk scut work but are now being edged out by an AI, I am wondering how those disciplines and corporations will be able to function in a decade’s time. Will they be forced to rely more and more on whatever version of AI is about then? Will smarter ones reach out to older staff for help in picking up the basics while they are still there? I can only imagine what sort of lawsuits will be arising out of the poor performance of those disciplines and corporations due to lack of competence. Will “schools” arise to teach the basics of what they missed to this generation? Don’t ask me why but I am reminded of how during the Vietnam war, US basic training got so bad that troops that arrived in Vietnam were sent to a second basic training course there to learn what they should have learnt in the states.
In the US at least, lawyers have to be certified as fit in terms of character for the bar by the ABA, plus there is a duty to do your own research and sign off that the research is yours when filing pleadings with the court. And further, you have to graduate from an ABA certified law school in most states. So unless there is some movement to get rid of the ABA as a certifying body, the legal system will be marked safe from AI, relatively speaking. Lower level paralegal types of jobs could be at risk and exposed to creeping enshittification, though.
Everyone likes to hate these gatekeepers, like the ABA, but they could end up saving us from dystopian outcomes.
To be fair there are some disturbing developments that could end up becoming a trend:
https://www.texasstandard.org/stories/texas-supreme-court-breaks-with-american-bar-association-over-law-school-credentials/
Sounds like the new business model for Google, MSFT, and Meta … make workers depend on AI to do the “thinking” for them, then once they’re totally dependent and unable to think for themselves, jack up the price.
Lawyers have already been filing AI research as ‘their own’ though, and for some reason continue to do so even after other lawyers were embarassingly exposed in court. (I don’t mean to say at all that this is all lawyers but I had thought that surely the first flop would have been caution enough)
The enshittification of goods and services is only half the story. This also would require maintaining an insufferable level of delusion and gullibility about the tech and “productivity.”
They need to tell themselves the lies about what the tech is supposed to cause.
More BS claims about the future from people that have no idea how to fix the problems of the present.
And what to make of this:
“Given that most of society is going to be unemployed (and unemployable), governments should be coming up with the 21st-century equivalent of a ‘new deal’, with social safety nets and training for this new world.
(Because all those AI-produced goods and services will need end-consumers with purchasing power.)”
Note the idea for “governments to provide training”. What? All this allegedly brilliant tech and industries can’t provide training?
Who is providing “training”? This is just something being made up because none of these “ideas” about the future pass the smell test.
The.increased profits won’t be because of improved goods and services, it will be because government would subsidize all the still needed labor costs. I imagine they are thinking of a way to do it without being called socialists.
The only other way to maintain this economic system that cannibalizes itself is to bring back slavery.
This is the most likely outcome…with alleged “AI” simply being the surveillance tools for management.
Then another part of me doesn’t want to engage with any of these hypotheticals because it’s ALL PART OF THE MARKETING HYPE. Gotta keep the bubble floating a bit longer.
“Après moi le déluge! Nach mir die Sintflut! Das ist der Wahlruf jedes Kapitalisten und jeder Kapitalistennation”
“Après moi le déluge! After me, the deluge! That is the battle cry of every capitalist and every capitalist nation.”
Karl Marx, Das Kapital I
And Louis XIV. One is tempted to say power absolutism is evergreen but it is more a deciduous plant–prone to dormancy in unfavorable climates.
Winter is coming….
Regarding AI and Cobol, one of the last big successful use case projects I worked on at the AI coding assistant product I am in the process of leaving was with a US organization to bring the AI coding assistant into the Cobol developers’ workflow along with a specific RAG connector to bring in context from their existing code to provide to the LLM (Claude, actually). It worked very well and they doubled their seat count with the company because of the results.
A possible reason why you don’t hear much about success stories of this nature are they are in the domain of the enterprise tools which when they do publish details about these types of successes they are in the form of sales marketing white papers that aren’t available for public distribution (this is the case with the project I worked on). Many of the most egregious horror stories come from publicity-generating projects, aka non-paying, so the details can be made public without the companies involved putting a stop to the publication when their business team gets wind of something that might reflect poorly on them through their involvement with a vendor.
Please note that I am not stating this is always or even the majority of cases with the negative stories you hear and that there is some huge well of hidden positive stories: I just have actual professional experience with the specific Cobol/AI coding assistant improvement use case that you mentioned. It worked well because there were use case-specific tools (the RAG connector, the IDE interface, a lot of privacy guarantees) between the user and the LLM.
Is it too early to claim success? My understanding is what is generated (Cobol to Java) is best described as Jobol and is difficult to maintain.
My admittedly old, but at that time up-to-date experience developing with COBOL, is that programming in that language itself is just one difficulty.
One also has to be versed in a variety of other software tools to build anything. In an IBM environment this was e.g. CICS, IMS or DB2, REXX (transaction manager, database management, scripting respectively), or in a DEC environment Datatrieve, RDB, DCL (report writing, database management, command language) — among others. I doubt that sort of constraint has disappeared (perhaps new software components have replaced legacy tools). A lot of COBOL code was actually generated from higher-level DDL, embedded SQL, and the like (the produced code was itself basically unreadable).
I handwaved through the description of the enterprise tooling because I am kind of weary of talking about this stuff but I am also weary of people not really getting the difference between a general purpose LLM thrown at a use case with it and with no tooling or RAG so let me explain more.
The IDE most developers use these days is VS Code. There are several others with significant market share but vsc is free and there is an open source version, so a lot of other IDE are forked from it. It’s so popular because it has a huge plugin marketplace that makes it trivial to expand with other tools like database connectors, test runners, device bus managers, linters, language server protocol implementations, whatever. Most AI coding assistants are either extensions (eg Cline) or forked versions of vsc (eg Cursor) so they can leverage the extensions. This would include all the Cobol-specific tooling you mention as well as a remote tunnel interface to the mainframe the cobol runs on.
The AI coding assistant itself is an application layer that connects to the LLM and provides additional tooling for interaction with the LLM and the IDE. Most important among those tools are those related to “context”, all the stuff passed to the LLM in addition to the prompt itself. This context can be the contents of open files in the text editor, functions explicitly referenced in the prompt, information called by a tool used by an agent or an intermediary application, or stuff on the RAG (the various indices that are formed by context connectors).
This additional tooling is the reason why AI coding assistants are one of the only true viable uses for the general purpose LLMs (go look at Cursor’s valuation and keep in mind it is not reliant on any single model and has fewer than 250 employees). It’s the reason a general purpose LLM can generate useful output not in it’s training set, which was the big breakthrough in the last 18 months, not the frontier models getting bigger.
I assess any potential commercial use case on the maturity of the tooling like I described for coding. Right now there aren’t any others. The closest on the horizon is vfx/animation/video editing and music production – these are industries that also use a software interface for their editing and production so like IDEs the tooling can be built into their existing workflow – but it will be constrained to things that aren’t under strong IP protection or enterprise companies leveraging their own IP to create the RAG pipelines to hyper niche models.
do keep posting. you’re one of a few commenting on here that are clearly day to day working with it. i.e. real life first hand experience.
and what an experience working with it is (if you put to one side the financial, water, energy and social problems … ).
I spent yesterday wrapping my head around skills in claude code, today working on a plugin for bitbucket, and the start of the week looking at data provenance and traceability ideas.
some of the crazier forward ideas are not in the labs. they are in the open source projects pushing what the labs have already built as far as it can go. any research paper published is based on 6-12 month old models (not state of the art)
I think we’re in for a wild ride. Djinn is not going back in the bottle.
It’s going to take a long time to absorb the amount of change released so far. let alone what we have in 1,2,3 years.
Sorry, can’t agree with enthusiastic views on this topic. Yes, simple tasks might be optimized, here we are often mixing verbosity of the code with it’s complexity. LLMs with all IDE supports work well for verbose code (big alpha), but at some level of abstraction (high beta), you clearly see where it hits its limits.
Situation with Cobol actually shows, that positive effect was achieved in the framework where complexity was in human perception and readability (still high alpha in play). And yes, if LLM was a newcomer into Cobol environment, positive shift will be emphasized first.
So, here is “but” section:
Algorithms which don’t exist yet. No sources, no “examples”, some traces in github, but nothing what actually fits. Here LLM becomes just “supportive”, not generative.
General problem solving, so, no references to possible solutions. LLM under-performs, because we don’t have tools, which returning “NaN”, if solution is not achievable.
You can clearly observe the reality checking on issue trackers and existing roadmaps for many of languages: the amount of reported issues is not getting lower, the amount of planned features is not growing. Which means, the effect is not observable yet, and we just enjoy low-hanging fruits.
Talking about video production, but keeping in mind, that we are programmers. Any error in the LLM output is a “gamma” part of complexity pie. It requires mental effort to recognize an error and make another attempt to fix it. Those back-and-forth cycles, is essentially what gamma describes. They are bearable to some level if we develop code, but cycling in video creation process makes even small errors painful. Thus, quality/productivity tradeoff is on the plate, because at some point, a human will be giving up. It’s not a big deal for mass media, but it’s lethal for productive coding.
At some point, the cost of prompt creation becomes equal to the cost of manual coding. And we have zero ideas how to deal with it, how to estimate and manage this in a balance at least.
a few points in response:
“enthusiastic views ” – my position is that the technology itself is much more useful than the naysayers say, and less game changing than the hype. financial, energy use, society impact needs urgently addressing and we don’t do that if we ignore where the tech is useful
“simple tasks might be optimized” – sounds like automation to me. hand write or get the machine to help write it quicker. but if it is a simple task then it is repeatable, known and a target for some bog standard code .
I am not entirely sure what you mean by alpha, beta and gamma
“Algorithms which don’t exist yet. ” – so what?
novel algorithmic work is a vanishingly small component of software engineering. we are reusing and recycling the same patterns that have existed for decades. if someone is not reusing then they are creating a support headache for themselves and team in future.
imagine the opposite – that some high percentage of algorithms were novel – in this world we would have a talent and training pipeline problem, and a maintenance problem.
“general problem solving” – I disagree. if there are solutions to a problem, access to websearch to research the nature of the problem, then gemini or latest opus are very good at suggesting solutions. but.. you need to be good at describing the problem in the first place, which means you inherently are closer to the solution anyway.
I have not really done any technical video work (well, once over a decade ago. had a PhD on the team that had spent 3 years looking at the same video clip . all 10 seconds of it. it made him slightly odd ;) ).
yes cycling in video will be painful. but video is not verifiable. Code is. it is why good teams spend so much time on the test suite, CI, CD etc. with video the human is the judge.
“At some point, the cost of prompt creation becomes equal to the cost of manual coding.” – this assumes that these systems don’t get better with time. and the last few years shows there is every incentive to make them better with time – see the performance of Cursor now vs at launch. or Claude code now versus anthropic models two years ago. and that is just the models themselves, not the usage, pattern libraries, scaffolding, fine tuning on local context etc that happens when we apply the models.
If I were to do a tldr for a coding assistant it is this:
“It’s like having a pair programming partner who’s already worked with every API and framework, even if they’re not an expert in your specific problem domain.”
and one final share.
This from Kent Beck of eXtreme Programming fame.
The Bet On Juniors Just Got Better
he makes the case that augmentation helps junior learn more quickly.
I wouldn’t categorize “naysayers” as a considerably large group. Those, who somehow biased by ethical, religious or other sort of concerns, and act fairly (so, they choose not to use any form of AI support) – those are a invisible in a society. All the others, who have concerns: statistical analysis and proper implementation is a strategy which is good enough for many of domains (traffic control, some management, some planning)
The term “simple task” also shifted significantly. Specifically it was used in a context of low-hanging fruits and successful Cobol implementation. What was considered “simple” before LLM, becomes trivial. “Simple” – is what LLM do well.
“alpha, beta, gamma” – was explained as a triplet right in the post.
“novel algorithmic work is a vanishingly small” – not in my domain. Any recursive iteration with multiple conditions should be emplaced, conditions understood – which is another problem for prompting. Parsing is hard, and never be easy. Deviations from common standards are everywhere, they are leaking from imported libraries, compiler updates.
https://github.com/antlr/antlr4/issues it’s just example. the amount of bugs in imported code is proportional to the library complexity and its value. In real circumstances, you just don’t have an opportunity to report a bug and expect it to be solved.
To add a little bit of cold water: https://youtrack.jetbrains.com/issues it shows one MILLION+ issues – it’s all over the infrastructure and JB products.
https://youtrack.jetbrains.com/issues/KT – iterates over 52 000 bugs and reports specifically for language we use.
#1 Any 3rd party library not only inherit problems, it creates a shadow area of what might be trusted.
What it implies for LLM use? It should be prompted with the whole set of those 52k issues; It’s millions of tokens, among which your “write me a function…” becomes a midget. Take in consideration #1 – technically, it implies, that all the source code from those libraries should be also included.
At this point we can see, that our LLM approach is leaky. Calculation-wise leaky. Too much input data and processing required to have code “clean”. In modern languages it calls “semantics”, “best practices”, “community standards” etc, and yet, it’s not guarantee a success, because it’s just a verbalization upon a swarm of intrinsic limitations.
The straightforward approach is to train LLM upon a specific version of the compiler – but that requires AGI to convert code into “issues”. You see where I’m going? To create a workable training set against each version of the compiler and all supported libraries we should use something smarter than we have. This is a bottleneck for the whole industry. Yes, AGI would fix all the issues at once in a reasonable time, but we are not there. And it’s also a question, if AGI might be achieved, because the quality of training data is low. This is the bottleneck I see, and brute force attempts, expressed in hundreds of billions dollars just show our inability to solve it. Scaling is the last resort.
You note: you are asking: “write a bash script” – and it never asks, which bash version we use, which version of linux utilities etc. That’s the point, where our prompting becomes harder and harder. Though, LLMs generally do a better job for shell (it’s well documented), for modern languages all what we have – is an Issue tracker.
So, prompting like: “It’s like having a pair programming partner who’s already worked with every API and framework, even if they’re not an expert in your specific problem domain.” seems to be not feasible. You can try – it won’t work. LLM interprets it as a prompt to be supportive and friendly and overconfident upon suggestions, which implies: it will be persisting on wrong solutions. Here you must be precise: library versions, imports, project structure, but it’s not guarantee you’ll be heard, because training data is ambiguous, companies rushed to report “success”, and midlevel devs wanted things to be easy.
You say,
The trend in the enterprise tooling for the coding assistants is creating context engines that manage all the associated rules, guidelines, standards etc that the prompts and agents need to adhere to. You’re making a lot of assumptions about how the tooling is working, and most of the assumptions are, again, operating off the public chatbot implementations and not the real tools. I don’t think it follows that every instance of using a library requires the library source code but in a lot of cases the connectors can be used to pull that in too.
From the suggestions you’re offering with issues it sounds like you have a perception of coding assistants being the same as automated or agentic issue resolution. I’m talking about the former. I cannot speak to the latter other than to say agents aren’t ready for prime time except in a very few areas, like multi-turn chat interactions based on a coding prompt (“reference #remote_repo_on_rag implementation and @existing_data_sheet_in_local_workspace and create a new C application using the same structure as the remote reference implemention but using the new sensor in the datasheet in the local workspace”) while you go do something else. Here the term ‘agent’ is different still from what you’re describing, because it’s still in the IDE, it’s not operating as a background process. Background agents are in some of the IDE assistants or being built out in the context engines. But they’re not what you appear to think they are.
But being precise is exactly how to make this work properly, along with building out the proper context, rules and guides. Those rules and guides and context engines are for exactly what you’re claiming: imports, library versions, internal project structure from context in the open workspace or from a connector on the RAG! Do you just add your junior devs to the repo and lock them in the cubicle and tell them not to come out until they have filed at least 5 PRs? I sure hope not, because it’s much more efficient to spend at least a few hours explaining the project structure and goals or at least pointing them to some documentation. The real coding assistants with the real enterprise tooling are the same way.
>> The real coding assistants with the real enterprise tooling are the same way.
Nope))). Because the amount of issues is growing in critical repositories. Of course, we can ignore issue tracker links, but all those companies use agents as well. And the amount of issues, reported monthly, is growing. JB is aware how to use Capilot, even better – they have their own product. All what you need in a one bucket: Your own Language, IDE, tooling pipelines, your proprietary training data, your model, but NOPE, even close: issue tracker might be only moderated.
https://www.jetbrains.com/ai/
Here what they say:
>> Powered by our proprietary models and best-in-class AI, our solutions are responsibly designed to ensure transparency, protect your privacy, and keep you in control.
Fix your language, first, JB ))) ha ha. That would reduce burden upon other systems.
So, from mine perspectives, I gave a try to coding assistance, using different tools, and never was impressed. Currently I’m implementing local server with pre-trained model, probably, one of deep-seek ones. No super expectations, but it can work during night time, doing simple requests.
>> Those rules and guides and context engines are for exactly what you’re claiming: imports, library versions, internal project structure from context in the open workspace or from a connector on the RAG!
This is a believe system. When developers report, that data is being obfuscated or ignored even in well-tailored setups, it’s an important signal. So, no, thanks))) What we need to realize: any LLM is being trained upon data which is at least 6 month old. It can’t properly incorporate fresh changes, which requires prompting to be exhaustive.
Again. There is no way to see coding assistance working without checking on industry health: Issue trackers, github PRs etc. I don’t see a significant progress there.
https://github.com/gradle/gradle here you can see, that in Dec 2025 was reported ~75 issues, compared to Jun 2024 it gives increase ~40%. So, how AI era impacted gradle as a company in 2025?
This is my “thermometer”. You are measuring average temperature in the whole hospital, I do the same in the morgue.
So, how industrial giants, whose performance degraded after AI era start, can demonstrate success of the whole AI project?
https://www.reuters.com/business/business-leaders-agree-ai-is-future-they-just-wish-it-worked-right-now-2025-12-16/
To be more precise, why prompting becomes a significant problem, and how it reflects in business applications, here just some picks from that link:
>> Part of the solution was designing prompts that gave the model permission to say no.
>> But Cando ran into a surprising stumbling block: the models couldn’t consistently and correctly summarize the Canadian Rail Operating Rules, a roughly 100-page document that lays out the safety standards for the industry.
>> AI researchers say models often struggle to recall what appears in the middle of a long document.
>> Many financial firms rely on data compiled from a broad range of sources, all of which can be formatted very differently. These differences might prompt an AI tool to “read patterns that don’t exist,” said Clark Shafer, director at advisory firm Alpha Financial Markets Consulting.
>> “People thought AI was magic. It’s not magic,” Beinat said. “There’s a lot of knowledge that needs to be encoded in these tools to work well.”
so, what I’m saying: general public believes, that coding is easy for LLMs. Nope. We just don’t report unsuccessful cases, code manually, when tasks are not trivial. From some perspectives, this coding applicability myth, is what actually drives the AI bubble.
a 100 page document is a huge amount of context for any but the most recent class of frontier models and at the time this testing was performed the available models simply would not have had the context window to handle loading it into a single prompt.
There is some contradiction I assume. 1M+ lines in a project is also a heavy burden. LLMs are not doing well even upon 100k lines of code.
a 100 page document, btw, is 4k lines of text. So we are comparing 100k lines of code with 4k of readable text.
In other words: when we are throwing some ideas into Copilot, which part of our codebase 100k will be ignored?
neat very specific project. I can imagine it very much giving a boost to someone that wants to take on Cobol work.
and… therein lies the rub.
people have to want to do it.
that can be solved of course – pay above the odds, actually take training seriously, plan properly and make a multi year commit to the team taking the unsexy cobol work on…
and people have to accept the scrutiny and risk. i saw the team at a major uk retailer – all near retirement a decade ago, and the mainframe still needed for the operations of every single one of the hundreds of stores in the UK. there were nine of them, with the “junior” being in his 40s. and a team in india being hurriedly trained to take over.
old systems like this are critical to many big businesses. they are old because they are successful (survivor effect). they have engineers in them happy to maintain something for decades rather than rush to the next new thing. or too expensive/risky to replace. which makes them deeply embedded in the overall architecture of many business.
so you need to be happy to not just take on the cobol, but the high profile project that replaces something that is working, that is integral to the business, in a language with no future. mmmmm. no.
I no longer read any news online. I have newspapers delivered to the house. This simple trip back to my youth has done wonders for my mental health. Interestingly, it is clear that many, although nowhere near most, newspaper articles are clearly being written by AI…..the same sing-song voice as online. That is a story for another day.
About 2 weeks ago in the papers, there were big discussions about specifically Meta, and how Zuckerberg was really abandoning the entire “metaverse” concept of virtual reality. The Zuck was so arrogantly confident about this that he renamed the company from Facebook to Meta. The uptake has been absymal. There were multiple reasons discussed but the most prominent was that more than half the users got physically ill after the first use and never went back again. It was apparently very difficult to retain users. I myself on a vacation a few years ago tried the goggles in an arcade with the kids. Within seconds of taking the goggles off and joining the real world ( it was some kind of game of traveling through a forestscape) I became violently ill with severe vertigo and a pounding headache and a vague sense of nausea that lasted the entire day. I had to go back to the hotel. Why would I ever touch that technology again? I understand from the articles I read that this is a real problem and with current technology there seems to be no simple and consistent way to fix this for most users. It is a game ender for more than half the users on their very first attempt.
So, that “next big thing”, virtual reality, got tanked because it was making the users physically ill. If only we would be so lucky that human users would notice that AI is making them spiritually, socially and intellectually ill. My opinion is that this is far worse than the physical illness of the metaverse. Humans are very in tune with physical illness, not so much the illness of the soul and intellect. And this is what scares me to death about this whole AI situation.
I had no idea that virtual reality was literally nausea-inducing for so many. Thanks for the report.
I would never try because I don’t have binocular vison so any attempt to create 3D would immediately give me a headache.
As someone without binocular vision (and hence no depth sight) 3D glasses and VR work about the same as a computer screen for me, except the eye and motion tracking allows the screen/glasses to look like a portal of sorts into the computer generated world.
I don’t get nausea or vertigo from it, either, which may be a personal quirk, though I suspect that how the binocular vision is fooled into believing an image very close to the eyes isn’t might be one reason for the deleterious effect.
Vertigo while using 3D is usually due to how movement is done; many interfaces use a controller that emulates flowing movement, as is common in regular video games, whereas another method is becoming common in VR applications where you teleport to another spot bypasses the “sight is moving without the body experiencing it” issue that might be another cause of problems.
Personally, I’ve found VR to be most useful in things like flight and car simulators, where your viewpoint is in a vehicle that moves. This likely uses our experience of cars to make sense of things.
By no means should this be read as a “you’re doing VR wrong” statement – but if you happen to try it out, I hope this helps.
PS.
For my binocular friends, using VR with just one eye may approximate my experience (it will completely remove the 3D effect), with the added advantage of looking like a pirate. YMMV.
DS
Yes, the physical nausea is quite common from what I’ve seen (as a retired technology person I’ve been around quite a few early adopters).
The VR idea has been tried several times in the past couple of decades, always with the same result. This is a paradigm for software development in general: failure doesn’t count because “now it’s different”. Software Sisyphus
same here.
I tried first and second versions of Oculus Quest.
Utterly and totally wonderfully engaging and immersive for a small number of games. A light saber type thing set to music – like a mash up of Star Wars and Just Dance. A boxing game that was a real work out.
and the party favourite – “The Plank” –
simple premise – walk into an elevator, go up to the 15th floor, the elevator doors open on a plank that is suspended above the ground 15 floors below. There are sound effects as you get closer, a helicopter in the distance. It is all very realistic.
We had a work christmas do and dialled it up to 11. I took an actual plank in, with a cushion under each end it gave an unstable floor (another sense fooled). and we had a large indoor fan to mimic the updraft.
with that combination 1 in 3 never made it out of the elevator at all. it just felt too realistic. some walked out part way then retreated to the lift. and only a very few managed to jump off the plank.
and if you jump off? you fall. very realistically, and most people on hitting the floor their knees would buckle and they’d fall to the floor.
extraordinary.
however… I describe that as the one party trick it had. that plus a few 3d games. otherwise it was useless. Too heavy on the head. and an hour or more of use would see me or my kids feeling sick from it. use it longer and risk a splitting headache.
we still have them, but they’ve not even been charged in 2025.
and as for work? forget it. too slow an interface. a solution looking for a problem.
and on the eyesight thing – there are calibration steps at the start so the quest can accommodate a wide range of different eye prescriptions. so it was good enough for short use
Thanks for the piece, which unfortunately is very depressing. My impression, founded on not particularly wide-ranging anecdata, is that AI is essentially a very polarizing tool: people either utterly adore it (both as work-tool and as companion) or refuse to engage with it in principle. The former, of course, need no convincing in order to use the tools, and can be relied upon by the corporations; the matter that needs work on is how to force the people who very much dislike it to use it all the same. Conor shared a link today about OpenAI getting into the pornography business, which seems like a way to expand markets, I suppose. But if we’re getting cases of psychosis from the current uses of the technology, I can only imagine how bad it’ll get when sexual content is actively promoted by the biggest AI corporation.
Porn drives pretty much all emerging interactive tech. No reason why AI will be different.
In links today:
Why Sam Altman is getting into the pornography business – Oligarch Watch
Over 35% of all web data transferred on the internet is dedicated to pornography
I’m thinking data-collection from users would alone be worth a fortune.
Bonus: each user gets their own, personalized Epstein file.
“Given that most of society is going to be unemployed (and unemployable), governments should be coming up with the 21st-century equivalent of a ‘new deal’, with social safety nets and training for this new world.”
The same government where the wrecking crews have set up shop for decades now?
These are some of the same people trying to create rifts to get rid of all the New Deal safety nets now saying there will be a government needed to provide safety nets.
Talk about speaking with a forked tongue.
If AI was rolled out like a normal technology that gets adopted in a more or less organic way, I doubt there would be much an issue with it being gradually adopted widely. But since it has been literally shoved down our throats in a rather openly forced uptake, we are where we are. I think Elon Musk is rather emblematic of who we are allowing to do this to us: “the fundamental weakness of Western civilization is empathy.” Since this is the very definition of a psychopath, I think it can be said that, while AI might be creating monsters, without a question it was created by monsters.
All the research says that children learn morality from experiencing how their parents treat them, and observing how parents treat others.
So in families where the parents treat their children badly and abusively, the odds of the children growing up the same is high. Not only do offspring have warped and selfish morals, but they have learnt ‘moral disengagement’ so that when they want something, they have ways of disabling their good morals with rationalisations and excuses.
So if parents, for whatever reason, are allowing chatbots to take over the moral education of their children, then its game over for socially cohesive morality.
And it appears to me that AI has scraped the internet and found the most manipulative, psychopathic, machiavelllian and sadistic ways of interacting with human beings. So I guess it copies its bro creators.
The world is turning into a late male adolescent, bro, instantiation.