No.

The bottleneck isn't intellectual productivity. The bottleneck is a legion of other things; regulation, IP law, marketing, etc. The executive email writers and meeting attenders have a swarm of business considerations ricocheting around in their heads in eternal battle with each other. It takes a lot of supposedly brilliant thinking to safely monetize all the things, and many of the factors involved are not manifest in written form anywhere, often for legal reasons.

One place where AI is being disruptive is research: where researchers are applying models in novel ways and making legitimate advances in math, medicine and other fields. Another is art "creatives": graphic artists and other. They're early victims and likely to be fully supplanted in the near future. A little further on and it'll be writers, actors, etc.

In the scenario being discussed - if a bunch of companies hired a whole bunch of lawyers, markerters, etc that might make salaries go up due to increased demand (but probably not super high amoung as tech isnt the only industry in the world). That still first requires companies to be hiring more of these types of people for that effect to happen, so we should still see some of the increased output even if there is a limiting factor. We would also notice the salaray of those professions going up, which so far hasn't happened.

In your own words, a business will run out of business quickly if supply and demand do not match. So unless you are confident that there will be a buyer, you cannot raise prices infinitely.

» because a living space is a basic necessity

While most people can live without a netflix subscription (hence they cannot raise prices infinitely and still expect to find buyers) most people prefer to live in housings. A housing is a basic necessity hence as a landlord you can confidently raise prices to the barrier of an affordability limit.

» Something else will get expensive in the meantime

Lets assume electricity prices get really cheap because humanity discovers fusion reaction. Well guess what, now landlords will increase the rents again because they can.

Hope I expressed myself to your liking. I mean you just walz in here and start lecturing people about capitalism, maybe you should change your careerpath and become a teacher.

The tech is going to have to be absolutely flawless, otherwise the uncanny-valley nature of AI "actors" in a movie will be as annoying as when the audio and video aren't perfectly synced in a stream. At least that's how I see it..

For most of them I'm not seeing any of those issues.

A couple years ago, we thought the trend was without limits - a five second video would turn into a five minute video, and keep going from there. But now I wonder if perhaps there are built in limits to how far things can go without having a data center with a billion Nvidia cards and a dozen nuclear reactors serving them power.

Again, I don't know the limits, but we've seen in the last year some sudden walls pop up that change our sense of the trajectory down to something less "the future is just ten months away."

The quick cuts thing is a huge turnoff so if they have a 15 second clip later on, I missed it.

When I say "1second" I mean that's what I was doing with automatic1111 a couple years ago. And every video I've seen is the same 30-60 generated frames...

Can you give an example, say in Medicine, where AI made a significant advancement? That is we are talking neural networks and up (ie: LLM) and not some local optimization.

"Our study suggests that LLMs have achieved superhuman performance on general medical diagnostic and management reasoning"

    > One place where AI is being disruptive is research: where researchers are applying models in novel ways and making legitimate advances in math, medicine and other fields.

    > In 2024, Hassabis and John M. Jumper were jointly awarded the Nobel Prize in Chemistry for their AI research contributions for protein structure prediction.

Related: (Also from Wiki):

    > Isomorphic Labs Limited is a London-based company which uses artificial intelligence for drug discovery. Isomorphic Labs was founded by Demis Hassabis, who is the CEO.

Yes, it's an example of ML used in science (other examples include NN based force fields for molecule dynamics simulations and meteorological models) - but a biologist or meteorologist usually cares little how the software package they are using works (excluding the knowledge of different limitation of numerical vs statistical models).

The whole thing "but look AI in science" seem to me like Motte-and-bailey argument to imply the use of AGI-like MLLM agents that perform independent research - currently a much less successful approach.

I specifically didn't call LLMs a statistical model - while they technically are, it's obvious they are something more. While intelligence is a hard concept to pin down, current gen LLMs already can do most (knowledge work) based things better than most people (they are better writers than most people, they can program better than most people, they are better at math than most people, have better medical knowledge than most people...). If the human is the mark of intelligence - it has been achieved.

Alphafold is something else though. I work with something similar (specifically FNOs for biophysical simulations) and the insight that data only models perform better than physics based model is novel - I think that the Nobel prize was deservedly awarded - however the thing is still closer to a curve fit than to LLMs regarding intelligence - or in other words, it's about as "intelligence" as permutation based black boxes were.

LLMs only exist because the companies developing them are so ridiculously powerful that can completely ignore the rule of law, or if necessary even change it (as they are currently trying to do here in Europe).

Remember we are talking about a technology created by torrenting 82 TB of pirated books, and that's just one single example.

"Steal all the users, steal all the music" and then lawyer up, as Eric Schmidt said at Stanford a few months ago.

They want to ban states from imposing their own regulations on AI.

They have trouble with debugging obvious bugs though.

Like let's take operating systems as an example. If there are great productivity gains from LLMs while aren't companies like Apple, Google and MS shipping operating systems with vastly less bugs and cleaning up backlogged user feature requests?

I don't get it either. You hire someone in the hope for ROI. Some things work some kinda don't. Now people will be n times more productive therefore you should hire fewer people??

That would mean you have no ideas. It says nothing about the potential.

Doesnt really matter if AI actually works or not.

It also matters a bit where the reputation cost hits. Layoffs can spook investors because it makes it look like the company is doing poorly. If the reputation hit for ai is to non-investors, then it probably matters less.

I think the reality is less like a switch and more like there are just certain jobs that get easier and you just need fewer people overall.

And you DO see companies laying off people in large numbers fairly regularly.

Sure but, so far, too regularly to be AI-gains-driven (at least in software). We have some data on software job postings and the job apocalypse, and corresponding layoffs, coincided with the end of ultra-low interest rates. If AI had a recent effect this year or last, its quite tiny in comparison.

https://fred.stlouisfed.org/graph/?g=1JmOr

so one can argue more is to come, but its hard to see how its had a real effect on jobs/layoffs thus far.

Shipping features faster != innovation or improvements to existing products

I’m not as bullish as some are on the impact of AI, but it does feel nice when you can deliver something in a fraction of the time it used to take. For me, it’s more useful as a research and idea exploration tool, less so about writing code. Part of that is that I’m in Scala land, so it just tends to not work as well as a more mainstream language.

We haven’t used it to help the product management and solution exploration side, which seems to be a big constraint on our execution.

What makes you so sure of the productivity boost when we aren't seeing a change in output?

LLMs are also not very useful for long term strategy or to come up with novel features or combinations of features. They also are not great at maintaining existing code, particularly without comprehensive test suites. They are good at coming up with tests for boiler plate code, but not really for high-level features.

From my experience, this stuff is rarely introduced to save developers from typing in the code for their logic. Reasons I observe:

1. SaaS sales/marketing pushing their offerings on decision makers - software being a pop culture, this works pretty well. It can be hard for internal staff to push back on What Everyone Is Using (TM). Even if it makes little to no sense.

2. Outsourcing liability, maintenance, and general "having to think about it". Can be entirely valid, but often it comes from an "I don't want to think of it" kind of place.

I don't see this stuff slowing down GenAI or not. Modern software is not unlike modern cars: A potpourri of components from different vendors that may all have been tested in isolation, but a lot of errors will arise from integration.

How do you know this? What are the bottlenecks?

One obvious one is: information gathering, processing, and comprehending.

How do you know what software to build without gathering, cleaning and manipulating, and finally understanding the "bigger picture" of data?

Also, sales & marketing. You would need to ramp up sales and marketing budgets/teams/training/so on to now push these products. More headcount and so forth.

Tbh, still a people problem. Software/devs are just tools to be used. It's not the tools themselves that are the bottleneck, but the ones using them.

/* -------------------- End Response ------------------------------------ */

IMO, we've hit a limit. What more tech is there to sell besides LLM integrations? The gen ai labs/foundational model researchers have already gotten so far ahead no one will be able to beat them (sans stealing/reverse engineering and reselling a la deepseek et al.).

What about old tech? Most of it is in maintenance mode. All the heavy lifting has been done by people with more skillset. Usually in these cases you downsize/offshore/cut costs to less technically skilled "janitorial teams." Perhaps some additions/expansion, but not significant enough to be more than a blip.

Really cannot think of anything to justify the amount of tech talent available globally or even domestically. What work is there for them to do lol? Software as an industry has really hit a ceiling -- I can't think of anything in recent times outside of LLMs that have provided outsized value (real or imagined).

Anyway! I left the industry after grinding to staff eng. Now finance/biz focused. Pure tech as an era is dead.

The company that I work for is currently innovating very fast (not LLM related), creating so much value for other companies that they have never gotten from any other business.. I know this because when they switch to our company, they tell us how much better our software product is compared to anything they've ever used. It has tons of features that no other company has. That's all I can say without doxxing too much.

I feel like it's unimaginative to say:

> What more tech is there to sell besides LLM integrations?

I have like 7 startup ideas written down in my notes app for software products that I wish I had in my life, but don't have time to work on, and can't find anything that exists for it. There is so much left to create

Now, there come a few considerations I don't believe you have factored in:

- Just because your company has struck gold: does that mean that pathway is available or realistic enough for everyone else; and to a more important point, is it /significant/ enough that it can scoop the enormous amount of tech talent on the market currently and in the future? I don't believe so.

- Segueing, "software products that I wish I had in my life." Yes, I too have many ideas, BUT: is the market (the TAM if you will) significant enough to warrant it? Ok, maybe it is -- how will you solve for distribution? Fulfillment is easy, but how are you going to not only identify prospective customers (your ICP), find them and communicate to them, and then convince them to buy your product, AND do this at scale, AND do this with low enough churn/CAC and high enough retention/CLTV, AND is this the most productive and profitable use of your time and resources?

Again, ideas are easy -- we all have them. But the execution is difficult. In the SaaS/tech space, people are burned out from software. Everyone is shilling their vibe-coded SaaS or latest app. That market is saturated, people don't care. Consumer economy is suffering right now due to the overall economy and so on. Next avenue is enterprise/B2B -- cool, still issues: buyer fatigue; economic uncertainty leading to anemic budgets and paralysis while the "fog" clears. No one is buying -- unless you can guarantee they can make money or you can "weather the storm" (see: AI, and all the top-down AI mandates every single PE co and board is shoving down exec teams throats).

I'm talking in very broad strokes on the most impactful things. Yes, there is much to create -- but who is going to create it and who is going to buy it (with what money?). This is a people problem, not a tech problem. I'm specifically talking about: "what more tech is there to sell -- that PEOPLE WILL BUY -- besides LLM integrations?" Again, I see nothing -- so I have pivoted towards finance and selling money. Money will not go out of fashion for a while (because people need it for the foreseeable future).

Ask yourself, if you were fired right now at this moment: how easy would it be for you to get another job? Quite difficult unless you find yourself lucky enough to have a network of people that work in businesses that are selling things that people are buying. Otherwise, good luck. You would have more luck consulting -- there are many many many "niche" products and projects that need to be done on small scales, that require good tech talent, but have no hope of being productized or scaled (hint!).

I do think I may struggle a bit to find something comparable to my current company, but we’re also hiring right now. And it’s a very small company in the grand scheme of things, even though we have customers much bigger.

I guess having that experience makes me think that there must be a lot of other small companies working in their own interesting niche, providing a valuable product for a subset of major companies. You just don’t usually know they exist unless you need their specific niche.

But I recognize your points too. It seems like the B-to-C space is really tricky right now, and likely fits closer with what you’re describing.

I think that the flip side is that a company doesn’t need to make it big to be successful. If you can hire 5 developers and bring in $2m/yr, there’s nothing at all wrong with that as a business. Maybe we will get lucky and the market will trend towards more of those to fill in the void that you mentioned. I think it could lead to a lot of innovation and a really healthy tech world! But maybe it’s just being overly optimistic to think that might be the path forward :)

Luckily software companies are not ball bearings factories.

Why wouldn't you just 10x the productive output instead?

If the competition instead uses their productivity boost to do layoffs and increase short term profits, you are likely to outcompete them over time.

It may help you build a real product feature quicker, but AI is not necessarily doing the research and product design which is probably the bottleneck for seeing real impact.

Maybe overall complexity creeping up rolls over any small gains, or devs are becoming more lazy and just copy paste llms output without a serious look at it?

My company didnt even adapt or allow use of llms in any way for anything so far (private client data security is more important than any productivity gains, which anyway seems questionable when looking around.. and serious data breaches can end up with fines in hundreds of millions ballpark easily).

Having worked on software infrastructure, it’s a thankless job. You’re most heroic work has little visibility and the result is that nothing catastrophic happened.

So maybe products will have better reliability and fewer bugs? And we all know there’s crappy software that makes tons of money, so there isn’t necessarily a strong correlation.

  > shipping features and fixes faster than ever before

Seriously, this shit can be solved with a regex...

The number of issues like these I see is growing exponentially, not decreasing. I don't think it's AI though, because it started before that. I think these companies are just overfitting whatever silly metrics they have decided are best

Firstly, the capex is currently too high for all but the few.

This is a rather obvious statement, sure. But the impact is a lot of companies "have tried language models and they didn't work", and the capex is laughable.

Secondly, there's a corporate paralysis over AI.

I received a panicky policy statement written in legalaise forbidding employees from using LLMs in any form. Written both out of a panic regarding intellectual property leaking but also a panic about how to manage and control staff moving forward.

I think a lot of corporates still clutch at this view that AI will push the workforce costs down and are secretly wasting a lot money failing at this.

The waste is extraordinary, but it's other peoples money (it's actually the shareholders money) and it's seen as being all for a good cause and not something to discuss after it's gone. I can never get it discussed.

Meanwhile, at a grass roots level, I see AI is being embraced and is improving productivity, every second IT worker is using it, it's just that because of this corporate panicking and mismanagement, it's value is not yet measured.

> This is a rather obvious statement,

Nobody is saying companies have to make LLMs themselves.

SASS is a thing.

In regards to Private LLMs, the situation has become disappointing in the 6 months.

I can only think of Mistral as being a genuine vendor.

But given the limitations in context window size, fine tuning is still necessary, and even that requires capex that I rarely see.

But my comment comes from the fact that I heard from several sources, smart people say "we tried language models at work and it failed".

However in my discussion with them, they have no concept of the size of the datacentres used by the webscalers.

The tools are often cringe because the capex was laughable. E.g. one solution, the trial was done using public LLMs and then they switched over to an internally built LLM which is terrible.

Or, secondly, the process is often cringe because the corporate aims are laughable.

I've had an argument with a manager making a multi-million dollar investment in a zero coding solution that we ended up throwing in the bin years later.

They argued that they are going with this bad product because "they don't want to have to manage a team of developers".

They responded "this product costs millions of dollars, how dare you?"

How dare me indeed...

They promptly left the company but it took 5 years before it was finally canned, and plenty of people wasted 5 years of their career on a dead-end product.

The Google web-based office productivity suite is similar. I heard a rumor that at some point Google senior mgmt said that nearly all employees (excluding accounting) must use Google Docs. I am sure that they fixed a huge number of bugs plus added missing/blocking feature, which made the product much more competitive vs MSFT Office. Fifteen years ago, Google Docs was a curiosity -- an experiment for just how complex web apps could become. Today, Google Docs is the premiere choice for new small businesses. It is cheaper than MSFT Office, and "good enough".

Worker productivity is secondary to business destruction, which is the primary event we're really waiting for.

So let me keep it real, I am shorting Atlassian over the next 5 years. Asana is another, there's plenty of startup IPOs that need to be shorted to the ground basically.

I think that this sentiment, along with all of the hype around AI in general, is failing to grasp a lot of the complexity around software creation. I'm not just talking about writing the code for a new application - I'm talking about maintaining that application, ensuring that it executes reliably and correctly, thinking about the features and UX required to make it as frictionless as possible (and voice input isn't the solution there, I'm very confident of that).

I'll be here in a year, we can have this exact discussion again.

"AI" is not going to wholesale replace software development anytime soon, and certainly not within a year's time because of the reasons I mentioned. The way you worded your post made it sound like you believed that capability was already here - nevertheless, whether you think it's here now or will be here in a year, both estimates are way off IMO.

Me too. Mostly so I can laugh though.

In smaller businesses some roles won’t need to be hired anymore.

Meanwhile in big corps, some roles may transition from being the source of presumed expertise to being one neck to choke.

I’d love it not to be true, but the truth is Jira is to projects what Slack/Teams are to messaging. When everybody is a project manager Jira gets paid more, not less.

Realistically though, they might incorporate that high schooler's software into Jira, to make it even more bloated and they will sell it to your employer soon enough! Then team lead Chris will enter your birthday and your vacation days in it too, to enable it to also do vacation planning, without asking you. Next thing is, that Atlassian sells you out and you receive unsolicited AI calls for your holiday planning.

When I used a not-so-simple LLM to make it act as a text adventure game it could barely keep track of the items in my inventory, so TBH i am a little bit skeptical that an LLM can handle entire project management - even without voice.

Perhaps it might be able to use tools/MCP/RPC to call out to real project management software and pretend to be your accountant/manager/whoever, but i wouldn't call that the LLM itself doing the project management task - and someone would need to write that project management software.

We just have to wait for the cards to flip, and that’s happening on a quadratic curve (some say exponential).

Content producers are blocking scrapers of their sites to prevent AI companies from using their content. I would not assume that AI is either inevitable or on a easy path to adoption. AI certainly isn't very useful if what it "knows" is out of date.

https://www.ft.com/content/4f20fbb9-a10f-4a08-9a13-efa1b55dd...

    > The bank [Goldman Sachs] now has 11,000 engineers among its 46,000 employees, according to [CEO David] Solomon, and is using AI to help draft public filing documents.

    > The work of drafting an S1 — the initial registration prospectus for an IPO — might have taken a six-person team two weeks to complete, but it can now be 95 per cent done by AI in minutes, said Solomon.

    > “The last 5 per cent now matters because the rest is now a commodity,” he said.

Note: I’m talking about your run of the mill SE waggie work, not startups where your food is based on your output.

E.g. look at the indie games count on steam by year: https://steamdb.info/stats/releases/?tagid=492

For example, I founded a SaaS company late last year which has been growing very quickly. We are track to pass $1M ARR before the company's first birthday. We are fully bootstrapped, 100% founder owned. There are 2 of us. And we feel confident we could keep up this pace of growth for quite a while without hiring or taking capital. (Of course, there's an argument that we could accelerate our growth rate with more cash/human resources)

Early in my career, at different companies, we often solved capacity problems by hiring. But my cofounder and I have been able to turn to AI to help with this, and we keep finding double digit percentage productivity improvements without investing much upfront time. I don't think this would have been remotely possible when I started my career, or even just a few years ago when AI hadn't really started to take off.

So my theory as to why it doesn't appear to be "painfully obvious": you've never heard of most of the businesses getting the most value out of this technology, because they're all too small. On average, the companies we know about are large. It's very difficult for them to reinvent themselves on a dime to adapt to new technology - it takes a long time to steer a ship - so it will take a while. But small businesses like mine can change how we work today and realize the results tomorrow.

So then I start thinking ... what sort of things am I doing that take me away from talking to customers? I spend a lot of time on implementation. I spend a lot of time on administrative sales tasks (chasing people for meetings, writing proposals, negotiating contracts). I spend a lot of time on meeting prep and follow-up. And many more. So I'm always on the hunt for tools with a problem already in mind.

In terms of specific tools...

Claude is a great backbone for a lot. Both the chatbot but also the API. I use the chatbot to help me write proposals and review contracts. I used it to write scripting to automate our implementation process which was once quite manual and is now a button click.

Cursor has been a game changer. In particular, it means that we spend very little time on bugfixes and small features. This keeps my CTO almost 100% focused on big picture needle-moving projects. We are now doing some research into things like Codex/Claude Code to see how we could improve this further.

Another app that I really love is called Granola. It automatically joins all of my meetings, writes notes, reminds me what promises I made, helps me write follow-up emails, and helps me prep for meetings.

Finally, we use an email client called Sedna (disclaimer: I used to work at Sedna) which is fully programmable. We've been building our own internal tooling (leveraging the Claude API) on top of Sedna to help automate different workflows. For example, my inbox is now perfectly prioritised. In many cases, when I receive emails from customers, an AI has already written a draft that I can review and send. I know there are a lot of out-of-the-box tools out there like Fyxer to help with things like this, but I've really appreciated the ability to get exactly what we want by building certain things ourselves.

Companies that needed to hire 10 people to grow, only need to hire 9 now

In less than 5 years that’s going to be 7 or 6 people

I’m doing more with 5 engineers than I was able to do with 15 just 10 years ago

Part of that is libraries etc have matured too but we’ve reached the point from a developer perspective that you don’t need to build new technologies, you just need to put what exists together in new ways

All the parts exist for any technology to be built, it’s about composition and distribution at this point

In 1987 the economist Robert Solow said "You can see the computer age everywhere but in the productivity statistics".

We should remark he said this long before the internet, web and mobile, so probably the remark needs an update.

However, I think it cuts through the salesmen hype. Anytime we see these kinds of claims we should reply "show me the numbers". I'll wait until economists make these big claims, will not trust CEOs and salesmen.

Only if you want to add "internet, web, and mobile" before "age". Otherwise it doesn't need any change.

But that phrase is about the productivity statistics, not about computers or actual productivity.

The problem with computers not changing the productivity statistics is one of the great mysteries economists argue about. It's very clear nowadays that there are problems on both the "statistics" and "productivity" sides of it, but the internet, web, and mobile didn't change anything.

Ok, so by 2027 we should be having fleets of autonomous AI agents swarming around every bug report and solving it x times faster than a human. Cool, so I guess by 2028 buggy software will be a thing of the past (for those companies that fully adopt AI of course). I'm so excited for a future where IT projects stop going overtime and overbudget and deliver more value than expected. Can you blame us for thinking this is too good to be true?

In complex systems, you can't necessarily perceive the result of large internal changes, especially not with the tiny amount of vibes sampling you're basing this on.

You really don't have the pulse on how fast the average company is shipping new code changes, and I don't see why you think you would know that. Shipping new public end-use features isn't even a good signal, it's a downstream product and a small fraction of software written.

It's like thinking you are picking up a vibe related to changes in how many immigrants are coming into the country month to month when you walk around the mall.

That doesn't mean it isn't a real productivity gain, but it might be spread across enough domains (bugs, features, internal tools, experiments) to not be immediately or "painfully obvious".

It'll probably get more obvious if we start to see uniquely productive small teams seeing success. A sort of "vibe-code wonder".

The more likely scenario is that if those tools make developer so much more productive, we would see a large surge in new companies, with 1 to 3 developers creating things that were deemed too hard for them to do.

But it's still possible that we didn't give people enough time yet.

AI also helps immensely in creating those other inefficiencies.

That said: it’s one type of work that is getting dramatically cheaper. The debate is about the scope and quality of that labor, not whether it’s cheap or fast (it is). But if anything negative (errors, faults) compound, and the correction can NOT be done with the same tools, then you must still have humans triage errors. In my experience, bad code can already have negative value (it costs more to fix than rewrite).

In the medium term, the actual scope and ability for different tasks will remain unknown. It takes a lot of time to gather the experience to tell if something was a bad idea – just look at the graveyard of design patterns, languages and software practices. Many of them enjoyed the spotlight for a decade before the fallout hit.

Anyway, while the abilities are unknown, AI will be used everywhere for everything – which is only wise if it’s truly better at every general task – despite every available data about it shows vastly different ability in different domains/problem types. Many of those things will be both (a) worse than humans and (b) expensive to reverse, with compounding effects.

The funny thing is I have already seen enthusiasts basically acknowledging this but explaining that those compounding issues (think tech debt) is the right choice now because better AI will fix those issues in the future. To me, this feels like the early formations of religion (not metaphorically even). And I have a feeling that the goalpost moving from both sides will lead to an unfalsifiability deadlock in the debate.

This eventually changed. Companies do figure out how to use tech, it just takes a while.