This reminds me of Antirez's "Don't fall into the anti-AI hype" [0]
In a sentence: These foundation models are really good at optimizing these extremely high level, extremely well defined problem spaces (ie multiply matrices faster). In Antirez's case, it's "make Redis faster".
There have been two reactions: "Oh it would never work for me" and "I have seen months of my life accomplished in an hour", and I think they're both right. I think we should be excited for Antirez, (who has since been popping off [1]), and I think the rest of us should rest easy knowing that LLM's can't (and maybe were never meant to) tackle the tacit-knowledge-filled, human-system-centric, ambiguously-defined-problem-space jobs most mortals work.
I have found Claude et al good at quickly implementing the algorithm I have in mind effectively, as long as I ask lots of control questions and check code. They aren’t good at inventing non-mainstream algorithms though and often slip staggeringly short term shortcuts in though. They are still a tool and not yet the craftsman who wields tools effectively. This will steadily change, and the corners where the obscure algorithm wins will erode further too.
>I think the rest of us should rest easy knowing that LLM's can't (and maybe were never meant to) tackle the tacit-knowledge-filled, human-system-centric, ambiguously-defined-problem-space jobs most mortals work
I don't believe that anymore, to be honest. Models are starting to get good at ambiguity. Claude Code now asks me when something is ambiguous. Soon, all meetings will be recorded, transcribed and stored in a well-indexed place for the agents to search when faced with ambiguity (free startup idea here!). If they can ask you now, they'll be able to search for the answers themselves once that's possible. In fact, they already do it now if you have a well-documented Notion/Confluence, it's just that nobody has.
It's probably harder to RL for "identify ambiguity" than RL'ing for performance algorithms, sure, but it's not impossible and it's in the works. It's just a matter of time now.
Unfortunately you can't record meetings in many jurisdictions, including court sessions. Hence we have to rely - for worse, or perhaps even for better - on human driven note taking.
You're downplaying the AI lobby here. They're eating down copyright laws, something that seemed impossible just a couple of years ago. Screwing privacy laws is just the next step.
> Soon, all meetings will be recorded, transcribed and stored in a well-indexed place for the agents to search when faced with ambiguity (free startup idea here!)
We were doing that over at Vowel a few years back, unfortunately it didn't pan out because you're competing directly against Zoom, Google Meet, Microsoft Teams, etc. They are all (slowly) catching up to where we were as a scrappy startup 4 years ago.
It was truly game-changing to have all of your meetings in an easily searchable database. Even as a human.
Private investment in the US has grown from 100 billion in 2024 to almost 300 billion USD in 2025 [0]. Add public investments worldwide and private investments in at least China and Europe.
I'm pretty sure money is not going to be the blocker.
Why not both? You don’t need 1trillion allocated before you have a proof of concept to demonstrate your non-LLM model, and once you have a PoC you will definitely have the larger investors interested
Advanced Machine Intelligence (AMI), a new Paris-based startup cofounded by Meta’s former chief AI scientist Yann LeCun, announced Monday it has raised more than $1 billion to develop AI world models.
LeCun argues that most human reasoning is grounded in the physical world, not language, and that AI world models are necessary to develop true human-level intelligence. “The idea that you’re going to extend the capabilities of LLMs [large language models] to the point that they’re going to have human-level intelligence is complete nonsense,” he said. [0]
I don't think it's valid to draw broad conclusions from the funding of a new company vs. an industry leader. If AMI builds something that looks impressive considering the funding they got, then they'll get plenty more in the next round.
Its not hard to tell at all, just look at how much it costs to run a 10T param model (especially with parallelized agents). Those costs are not worth the occasional slot machine-eque jackpot you get. For an entity like Google it might be worth it, but that's it. They definitely aren't going to let us use these things for cost they are now for much longer.
Imagine going back to 2020 and tell people in 6 years going to be able to spend $200.00 a month and be able to spin up $2mm in GPUs at full throttle to respond to your emails. None of this makes sense.
I don't know, I guess it depends from a) how many hours per month you spend answering emails, and b) how much more revenue you could get in that same time. $200 should be reasonably 2/3 hours of work? So that's about the amount of saved time per month to break even on your subscription. It's a steal.
Whenever you solve any hard problem, you start off by finding a complicated solution, which you then scale down to a simpler solution.
LLMs are a "complicated solution" in the sense that they're expensive. Once you know what they're capable of, you can scale them down to something less expensive. There's usually a way.
Also, an important advantage of LLMs over other approaches is that it's easy to improve them by finding better ways of prompting them. Those prompting strategies can then get hard-coded into the models to make them more efficient. Rinse and repeat. Similarly, you can produce curated data to make them better in certain areas like programming or mathematics.
>I think the rest of us should rest easy knowing that LLM's can't (and maybe were never meant to) tackle the tacit-knowledge-filled, human-system-centric, ambiguously-defined-problem-space jobs most mortals work.
A Statement all but guaranteed to look incredibly short sighted by 2030.
There is an apples and oranges difference between AI improving itself (becoming more capable) and AI optimizing software that happens to be used for AI training or inference.
A more efficient transformer just costs less to run.
"AI improving AI" would be if one generation of AI designed a next-gen AI that was fundamentally more capable (not just faster/cheaper) than itself. A reptilian brain that could autonomously design a mammalian brain.
Even when hooked up into a smart harness like AlphaEvolve, I don't think LLMs have the creativity to do this, unless the next-gen architecture is hiding in plain sight as an assemblage of parts than an LLM can be coaxed into predicting.
More likely it'll take a few more steps of human innovation, steps towards AGI, before we have an AI capable of autonomous innovation rather than just prompted mashup generation.
> Do we have other examples of AI being used to improve the LLMs
Yes, last year when they revealed AlphaEvolve they used a previous gemini model to improve kernels that were used in training this gen models, netting them a 1% faster training run. Not much, but still.
This is the thing to look for in 2027, imho. All the big AI labs have big projects working on research agents, also specifically into improving AI (duh) and I expect a lot of that to get out of the experimental phases this year.
Next year they actually get to do a lot of work and I think we will see the first big effective architectural change co-invented by AI.
Note that coding is not the only use of Gemini or any of these models. It's also not what this article is talking about. Gemini can be not the best coding agent, but very good at other things.
If you mean specifically the Gemini VS Code Extension: it's terrible compared to Claude Code or Codex. I don't know how they can get away with it. Just constant timeouts, weird failure modes, have to start a new chat to switch modes... but I don't think any of that is specific to gemini the model- it seems to be the extension.
As for actual solutions to problems ignoring the VS Code extension aspect, I find all three premiere models to be excellent coding agents for my purposes.
The overall quality of LLM coding tools is shockingly bad. I haven't found a single one without major issues, and many have the same problems reappear every few months, sometimes bad enough to almost break the entire thing (e.g. 100% failure rate in editing files, broken for weeks, with the same cause each time, multiple times in a year).
I for one can't tell the difference between Claude and Gemini for coding. And the internal agent tooling is many times faster than Claude Code in my experience.
> He says the problem is that they can't use Claude Code because it's the enemy, and Gemini has never been good enough to capture people's workflows like Claude has, so basically agentic coding just never really took off inside Google. They're all just plodding along, completely oblivious to what's happening out there right now.
This is a bunch of gabagoo. Wrong on so many layers, it's not even worth reading further.
a) goog has agentic coding in both antigravity & cli forms. While it is not at the level of cc + opus, it's still decent.
b) goog has their own versions of models trained on internal code
c) goog has claude in vertex, and most definitely can set it up in secure zones (like they can for their clients) so they'd be able to use claude (at cost) within their own projects.
I’m not so sure. From talking to some of my own friends at google they feel that antigravity/gemini models are handicapping them and would much rather be using claude code (which only deepmind gets to use)
I wish that Google would focus on bringing their Gemini 3.x models to GA, and provide enough capacity such that one not constantly has to fight with 429 errors.
It often feels like they do not want me to develop applications for corporate clients using their Vertex API. It is just such a shame, given that their models were so great for document analysis etc.
No, for clients we use paid Vertex AI accounts. We often need to host workloads in an EU region, which rules out “global” models (and probably better capacity).
In the past, we used a wrapper that round-robined across multiple projects to get enough quota. Luckily, many of our workloads are workflow-style tasks, so we can simply keep retrying on 429s.
Fun fact: for one of their services, I think it was Stitch, I noticed that my paid key kept hitting quota, while the free worked fine. That blew my mind.
How many times we have to hear again about Erdös problems? :) It sounds like a great achievement for humanity at first, but after a while they keep coming back!
I would be interested to see how exactly the agent helped. How was it used, where did it lead to the given improvement and in how far would it have taken a human to come to the same solution.
The CANOS arxiv link says absolutely nothing about AlphaEvolve, Gemini, or LLMs. It seems to use purely traditional ML models. If AE did in fact write a quick script to test different configurations in order to optimize the results, they don't seem to have bothered to write about it.
I can't read the Nature paper about DeepConsensus, but from the summary, it doesn't really explain what role AE had in improving DC. It would be nice to be able to read about what role it actually played, and whether it used traditional or novel methods of performing it
We went from 'AI will replace programmers' to 'AI will help programmers' to 'AI writes code while other AI reviews it' in about 18 months. At this rate the humans are just providing the electricity.
AlphaEvolve couples map-elites with LLMs. It's an key step in machine learning, in the vein of DQN for reinforcement learning.
AE brings diversity from the genetic algorithms community to large scale optmized deep learning and RL models.
It is a mandatory step for moving forward. The approach is clean and simple, while generic.
The only caveats is the per optimization problem definition of the map élites dimensions. But surely, this will get tackled somehow over the next few years.
If you don't know about map-elites, go look up Jean-Baptiste Mouret' s work and talks, it's both very interesting and universal.
From the comments it seems that this community (mostly career software people) is starting to move into a new phase of grief about the median software engineer losing their hoped for permanent place in society.
-2021-2024 was Denial
-2024-2025 was Anger and Bargaining
-2026 seems to be some combo of anger, bargaining and acceptance depending mostly on your class/age
What I'm most curious about is how this translates to messy, real-world codebases without well-defined metrics. Most production software isn't chip design or kernel optimization - it's business logic with unclear success criteria. The infrastructure story is impressive, but I'd love to see how they handle domains where the evaluation function itself is ambiguous.
> In advertising and marketing, WPP used AlphaEvolve to refine AI model components, navigating complex, high-dimensional campaign data and achieving 10% accuracy gains over their competitive manual model optimizations.
Ah good, we're getting closer and closer to Venus, Inc. every day. /s
This reminds me of Antirez's "Don't fall into the anti-AI hype" [0]
In a sentence: These foundation models are really good at optimizing these extremely high level, extremely well defined problem spaces (ie multiply matrices faster). In Antirez's case, it's "make Redis faster".
There have been two reactions: "Oh it would never work for me" and "I have seen months of my life accomplished in an hour", and I think they're both right. I think we should be excited for Antirez, (who has since been popping off [1]), and I think the rest of us should rest easy knowing that LLM's can't (and maybe were never meant to) tackle the tacit-knowledge-filled, human-system-centric, ambiguously-defined-problem-space jobs most mortals work.
[0] https://antirez.com/news/158 [1] https://antirez.com/news/164
I have found Claude et al good at quickly implementing the algorithm I have in mind effectively, as long as I ask lots of control questions and check code. They aren’t good at inventing non-mainstream algorithms though and often slip staggeringly short term shortcuts in though. They are still a tool and not yet the craftsman who wields tools effectively. This will steadily change, and the corners where the obscure algorithm wins will erode further too.
>I think the rest of us should rest easy knowing that LLM's can't (and maybe were never meant to) tackle the tacit-knowledge-filled, human-system-centric, ambiguously-defined-problem-space jobs most mortals work
I don't believe that anymore, to be honest. Models are starting to get good at ambiguity. Claude Code now asks me when something is ambiguous. Soon, all meetings will be recorded, transcribed and stored in a well-indexed place for the agents to search when faced with ambiguity (free startup idea here!). If they can ask you now, they'll be able to search for the answers themselves once that's possible. In fact, they already do it now if you have a well-documented Notion/Confluence, it's just that nobody has.
It's probably harder to RL for "identify ambiguity" than RL'ing for performance algorithms, sure, but it's not impossible and it's in the works. It's just a matter of time now.
Unfortunately you can't record meetings in many jurisdictions, including court sessions. Hence we have to rely - for worse, or perhaps even for better - on human driven note taking.
You're downplaying the AI lobby here. They're eating down copyright laws, something that seemed impossible just a couple of years ago. Screwing privacy laws is just the next step.
> Soon, all meetings will be recorded, transcribed and stored in a well-indexed place for the agents to search when faced with ambiguity (free startup idea here!)
We were doing that over at Vowel a few years back, unfortunately it didn't pan out because you're competing directly against Zoom, Google Meet, Microsoft Teams, etc. They are all (slowly) catching up to where we were as a scrappy startup 4 years ago.
It was truly game-changing to have all of your meetings in an easily searchable database. Even as a human.
So self chosen total surveillance and transparency so your fav LLM can be better?
Could always use a local LLM for stuff like that. One of my relatives works for one of the big audit firms and that's what they do.
In coding the ambiguity is very, very limited and constrained compared to any non dev job that involves any decision making
> I think the rest of us should rest easy knowing that LLM's can't [...]
What if (when?) (AI-assisted) research moves AI beyond LLMs? Do you think that can't happen?
Not in the next decade. Won't get funded.
Private investment in the US has grown from 100 billion in 2024 to almost 300 billion USD in 2025 [0]. Add public investments worldwide and private investments in at least China and Europe.
I'm pretty sure money is not going to be the blocker.
[0] https://hai.stanford.edu/ai-index/2026-ai-index-report
The money will go to LLMs.
Why not both? You don’t need 1trillion allocated before you have a proof of concept to demonstrate your non-LLM model, and once you have a PoC you will definitely have the larger investors interested
You will need 100s of billions to make a viable POC.
For a PoC? That sounds very unlikely. I think you’re off by at least 2–3 orders of magnitude
Advanced Machine Intelligence (AMI), a new Paris-based startup cofounded by Meta’s former chief AI scientist Yann LeCun, announced Monday it has raised more than $1 billion to develop AI world models.
LeCun argues that most human reasoning is grounded in the physical world, not language, and that AI world models are necessary to develop true human-level intelligence. “The idea that you’re going to extend the capabilities of LLMs [large language models] to the point that they’re going to have human-level intelligence is complete nonsense,” he said. [0]
[0] https://www.wired.com/story/yann-lecun-raises-dollar1-billio...
Now check how much OpenAI got in their last funding round, and you have your answer.
I don't think it's valid to draw broad conclusions from the funding of a new company vs. an industry leader. If AMI builds something that looks impressive considering the funding they got, then they'll get plenty more in the next round.
1B is what Microsoft invested in Open AI in 2019[0]. That was enough to get the ball rolling.
[0] https://en.wikipedia.org/wiki/OpenAI#Creation_of_for-profit_...
I'd say it's a malefactor of:
1. Amazing, you just tweaked 1% efficiency
2. You idiot, you just spent an hour trying to trouble shoot a hallucinated api.
On average, it's really hard to tell which ones going to win here.
Its not hard to tell at all, just look at how much it costs to run a 10T param model (especially with parallelized agents). Those costs are not worth the occasional slot machine-eque jackpot you get. For an entity like Google it might be worth it, but that's it. They definitely aren't going to let us use these things for cost they are now for much longer.
Imagine going back to 2020 and tell people in 6 years going to be able to spend $200.00 a month and be able to spin up $2mm in GPUs at full throttle to respond to your emails. None of this makes sense.
You don't pay for a £200 a month account to respond to your emails, and if you are, I would tell you that you're wasting your money.
I don't know, I guess it depends from a) how many hours per month you spend answering emails, and b) how much more revenue you could get in that same time. $200 should be reasonably 2/3 hours of work? So that's about the amount of saved time per month to break even on your subscription. It's a steal.
Whenever you solve any hard problem, you start off by finding a complicated solution, which you then scale down to a simpler solution.
LLMs are a "complicated solution" in the sense that they're expensive. Once you know what they're capable of, you can scale them down to something less expensive. There's usually a way.
Also, an important advantage of LLMs over other approaches is that it's easy to improve them by finding better ways of prompting them. Those prompting strategies can then get hard-coded into the models to make them more efficient. Rinse and repeat. Similarly, you can produce curated data to make them better in certain areas like programming or mathematics.
Do you realize you're fighting a strawman or do you actually think this is a compelling argument?
>I think the rest of us should rest easy knowing that LLM's can't (and maybe were never meant to) tackle the tacit-knowledge-filled, human-system-centric, ambiguously-defined-problem-space jobs most mortals work.
A Statement all but guaranteed to look incredibly short sighted by 2030.
AI improving itself (or at least the architecture it runs on), the singularity is near as they say.
Do we have other examples of AI being used to improve the LLMs, apart for the creation of synthetic data and the testing of the models?
There is an apples and oranges difference between AI improving itself (becoming more capable) and AI optimizing software that happens to be used for AI training or inference.
A more efficient transformer just costs less to run.
"AI improving AI" would be if one generation of AI designed a next-gen AI that was fundamentally more capable (not just faster/cheaper) than itself. A reptilian brain that could autonomously design a mammalian brain.
Even when hooked up into a smart harness like AlphaEvolve, I don't think LLMs have the creativity to do this, unless the next-gen architecture is hiding in plain sight as an assemblage of parts than an LLM can be coaxed into predicting.
More likely it'll take a few more steps of human innovation, steps towards AGI, before we have an AI capable of autonomous innovation rather than just prompted mashup generation.
> Do we have other examples of AI being used to improve the LLMs
Yes, last year when they revealed AlphaEvolve they used a previous gemini model to improve kernels that were used in training this gen models, netting them a 1% faster training run. Not much, but still.
I feel like the most viral lately is https://github.com/karpathy/autoresearch
Self improving, doesn’t necessarily imply singularity right?
There still could be hard constraints to make singularity intractable or just such a long time horizon it’s not practical right?
> AI improving itself
This is the thing to look for in 2027, imho. All the big AI labs have big projects working on research agents, also specifically into improving AI (duh) and I expect a lot of that to get out of the experimental phases this year.
Next year they actually get to do a lot of work and I think we will see the first big effective architectural change co-invented by AI.
And then on 2028 we will be selling ice cream at the beach.
Shameless plug: https://huggingface.co/spaces/smolagents/ml-intern
It’s a simple harness around Opus, but with tight integration to Hugging Face infra, so the agent can read papers, test code and launch experiments
What are the benchmarks for this, in terms of costs of computation and error; cost to converge?
Re: hyperparameter tuning and autoresearch: https://news.ycombinator.com/item?id=47444581
Parameter-free LLMs would be cool
The hard part about this is for every few 'WOW', there's a lineage of 'you dumbass'.
I mean, if you can create aharrness to filter these two, sure, singularity away; it's really hard to see how someones gonna do that.
Are Googlers themselves happy using Gemini coding agent instead of Claude Code or Codex? (no snark, I'm really asking)
Note that coding is not the only use of Gemini or any of these models. It's also not what this article is talking about. Gemini can be not the best coding agent, but very good at other things.
If you mean specifically the Gemini VS Code Extension: it's terrible compared to Claude Code or Codex. I don't know how they can get away with it. Just constant timeouts, weird failure modes, have to start a new chat to switch modes... but I don't think any of that is specific to gemini the model- it seems to be the extension.
As for actual solutions to problems ignoring the VS Code extension aspect, I find all three premiere models to be excellent coding agents for my purposes.
The overall quality of LLM coding tools is shockingly bad. I haven't found a single one without major issues, and many have the same problems reappear every few months, sometimes bad enough to almost break the entire thing (e.g. 100% failure rate in editing files, broken for weeks, with the same cause each time, multiple times in a year).
I'd say I'm surprised by it, but uh
The point of dogfooding is exactly that: if we're unhappy, we're the ones to improve.
the engineers using gemini have no control over deepmind
Antigravity comes to mind
I for one can't tell the difference between Claude and Gemini for coding. And the internal agent tooling is many times faster than Claude Code in my experience.
they use claude code at deepmind
Codex?
Last month, Steve Yegge suggested that they are not: https://xcancel.com/Steve_Yegge/status/2043747998740689171
> He says the problem is that they can't use Claude Code because it's the enemy, and Gemini has never been good enough to capture people's workflows like Claude has, so basically agentic coding just never really took off inside Google. They're all just plodding along, completely oblivious to what's happening out there right now.
This is a bunch of gabagoo. Wrong on so many layers, it's not even worth reading further.
a) goog has agentic coding in both antigravity & cli forms. While it is not at the level of cc + opus, it's still decent.
b) goog has their own versions of models trained on internal code
c) goog has claude in vertex, and most definitely can set it up in secure zones (like they can for their clients) so they'd be able to use claude (at cost) within their own projects.
Agreed, however imo there is def some problems unique to Google which is making the internal experience less than ideal.
Hoping they can figure it out sooner rather than later.
Demis Hassabis chimed in on that thread and called it what it is: clickbait.
I’m not so sure. From talking to some of my own friends at google they feel that antigravity/gemini models are handicapping them and would much rather be using claude code (which only deepmind gets to use)
Sure, but there's cavernous distance between "google = john deere" and "darn I have to use Gemini"
This couldn't be further from the truth
There is value in the "eating your own dog food".
If internal staff aren't happy with the tools they build, typically that should drive improvements to their own tools
I wish that Google would focus on bringing their Gemini 3.x models to GA, and provide enough capacity such that one not constantly has to fight with 429 errors.
It often feels like they do not want me to develop applications for corporate clients using their Vertex API. It is just such a shame, given that their models were so great for document analysis etc.
Are you doing it on a free plan? I noticed they serve way more 429s on the free plan.
No, for clients we use paid Vertex AI accounts. We often need to host workloads in an EU region, which rules out “global” models (and probably better capacity).
In the past, we used a wrapper that round-robined across multiple projects to get enough quota. Luckily, many of our workloads are workflow-style tasks, so we can simply keep retrying on 429s.
Fun fact: for one of their services, I think it was Stitch, I noticed that my paid key kept hitting quota, while the free worked fine. That blew my mind.
How many times we have to hear again about Erdös problems? :) It sounds like a great achievement for humanity at first, but after a while they keep coming back!
A fantastically simple solution to improving algorithms, I wish I had this years ago in activation engineering: https://blog.n.ichol.ai/llm-activation-engineering-an-easy-f...
How do I access AlphaEvolve?
This is just a flex post. Be a billion dollar company or get out.
I would be interested to see how exactly the agent helped. How was it used, where did it lead to the given improvement and in how far would it have taken a human to come to the same solution.
seems like `karpathy/autoresearch` on steroids
The blog post has many links to papers and preprints discussing this exact question.
The CANOS arxiv link says absolutely nothing about AlphaEvolve, Gemini, or LLMs. It seems to use purely traditional ML models. If AE did in fact write a quick script to test different configurations in order to optimize the results, they don't seem to have bothered to write about it.
I can't read the Nature paper about DeepConsensus, but from the summary, it doesn't really explain what role AE had in improving DC. It would be nice to be able to read about what role it actually played, and whether it used traditional or novel methods of performing it
We went from 'AI will replace programmers' to 'AI will help programmers' to 'AI writes code while other AI reviews it' in about 18 months. At this rate the humans are just providing the electricity.
AlphaEvolve couples map-elites with LLMs. It's an key step in machine learning, in the vein of DQN for reinforcement learning.
AE brings diversity from the genetic algorithms community to large scale optmized deep learning and RL models.
It is a mandatory step for moving forward. The approach is clean and simple, while generic.
The only caveats is the per optimization problem definition of the map élites dimensions. But surely, this will get tackled somehow over the next few years.
If you don't know about map-elites, go look up Jean-Baptiste Mouret' s work and talks, it's both very interesting and universal.
From the comments it seems that this community (mostly career software people) is starting to move into a new phase of grief about the median software engineer losing their hoped for permanent place in society.
-2021-2024 was Denial
-2024-2025 was Anger and Bargaining
-2026 seems to be some combo of anger, bargaining and acceptance depending mostly on your class/age
RSI is here on the hardware level and on software level. Sprinkle with a couple algorithmic breakthroughs and results are nigh unimaginable.
What I'm most curious about is how this translates to messy, real-world codebases without well-defined metrics. Most production software isn't chip design or kernel optimization - it's business logic with unclear success criteria. The infrastructure story is impressive, but I'd love to see how they handle domains where the evaluation function itself is ambiguous.
> In advertising and marketing, WPP used AlphaEvolve to refine AI model components, navigating complex, high-dimensional campaign data and achieving 10% accuracy gains over their competitive manual model optimizations.
Ah good, we're getting closer and closer to Venus, Inc. every day. /s