Ideally what I'd like to see is pluggable knowledge bases.
So if I'm e.g. coding a SwiftUI app for navigation, I'd take 9B of basic coding and reasoning, add 10B of swift/swiftUI, add 5B of GIS/geography knowledge and another 5B of frontend app design knowledge. My model doesn't need to know a single line of python.
Then when I want to research electronics components, I grab a 15B model of agentic research techniques, and add in 10B of electronics knowledge, etc.
I don't want general purpose models. They try to be everything to everyone. I want to click together a model that is laser-focused on what I am doing, and I want to run it locally
An LLM works better the more disparate world knowledge it has, even if it's not immediately obvious why it would be relevant. The model finds a structure to the problem you give it in a largely language-agnostic way that benefits from training on every language (these things are direct descendants of Google Translate), and even non-programming knowledge - the structure of your task might resemble an ancient Chinese poem that influences the model's response, for example. That structure is considered a form of compression, as some fascinating and illuminating recent 3blue1brown videos get into - a common pattern in Haskell or FORTRAN and a situation described in an ancient Chinese poem may all compress to something quite similar to your task, thus when the model compresses the idea of your task it immediately draws from those ideas.
There are "experts" which do divide parts of the model that are found to activate together for specific tasks, so they can be processed in parallel to join the result at the end, but it's nowhere near the granularity of a SwiftUI expert and a python expert. The difference in those things is so trivial from an abstract point of view that it would make no sense. They would be 99% the same.
Distillations also come into this but I'm highly skeptical you could make one guaranteed to only know programming and only in one programming language (especially with as small a sample set as SwiftUI relative to something like C) without its efficacy being hobbled by tunnel vision. Reminiscent of the SpongeBob episode where he empties his mind of everything except fine dining and breathing, then can't remember his name and goes insane. Beyond the basic concepts of general coding and the trivia of syntax, getting anything done requires a large intersection of disparate world knowledge and the ability to apply it to new situations.
This makes a ton of sense, and seems like it's coming from somebody who understands LLMs better than I do so I will definitely take your word for it - but it doesn't totally track with my experience of running small local models either. In my experience some models are better at language X and others better at language Y - and all of them are better at language X and Y than language Z-I-just-made-up. I take that to mean there is some relationship between training data and skill. Maybe not enough to make pluggable modules, tho...
Thanks for putting this so well. The mathematical evidence for this “general intelligence underlying everything” is the “universal geometry of embeddings” paper. Fascinating read. Or as the ancient philosophers used to say, the one who knows God knows everything.
This is a fundamental misunderstanding of how LLMs work. You can’t really specialize a model. You specialize the harness. A well-trained general purpose LLM doesn’t need examples in its training data, it can write good code in a new language you invented yesterday with just a spec definition. And it will perform better than a small model trained on lots of examples of your invented language. The reason is because of the “universal geometry of embeddings”, i.e all human languages have the same underlying pattern structure, so any model that is very good in any language is good in all languages. Attempting to specialize a model for a particular purpose often decreases overall performance. Fine-tuning is just a hack to make dumb models more reliable on limited tasks but they become incapable of doing anything else. Unless you are building a factory assembly line where a model is literally doing the same thing over and over, you almost always want a general purpose model over a specialized one.
This is so right. We training Whisper Large model on 20,000 audio samples specific to a domain and it ended up reducing the ASR by 5% while improving WER of the finetuned domain by 0.5%.
Instead we ended up with no finetuning. We give audio snippet to 2 AsR models, take 3 best transcriptions and ask the LLm to pick the best based on the context. That produced significantly higher accuracy in how an agent understands the users.
The problem with this idea is that knowing Python makes the model a better Swift programmer, as does a higher-number of parameters during training. So you'd be so much better off with a 90B general purpose model trained on everything anyway.
> I want to click together a model that is laser-focused on what I am doing
This is roughly what multi-agent systems are built for.
This is possible with models too, but "making one on the fly" is much easier with agent coordination rather than model weights, since they all speak the same language.
There is an IBM Mainframe vs Google Distributed system division here. Like Seymour Cray said - two oxen or 1024 chickens.
Chickens are harder to harness, so a lot of my work is in sled-dog territory for agent harnesses & command structures.
> This is roughly what multi-agent systems are built for.
I think I disagree. For some things, maybe that works - but think of a multi-agent system where one agent understands the code, and passes it off to the reasoning agent to figure out what the bug is. This system is going to suck. Because encoding enough info to figure out what the bug is would just be dumping every single line of the code.
So say agent 1 (reasoning) asks agent 2 (swift) to explain what is happening in File.swift. Anything agent 2 passes to agent 1 short of the entire code is a lossy transfer - and then the bug gets missed.
Sounds like MoE but more pluggable. Not sure if anyone is researching something like this. I still think your swift or GIS model will need basic reasoning and coding to work, so it's more like multiple smaller models which you can load as needed, e.g. sub-agents for GIS needs the GIS model.
Another approach would be to have basic coding and reasoning model and then load specification for language and libraries into context, it could work for self-hosted models, but I don't want whole specification of the language to be send to API and waste tokens on that.
My motivation is somehow the model should be able to work without hitting an external resource every single time and somehow have all the knowledge necessary locally so we don't get rate limited but how can you stay up to date with the latest information while still remaining a good boy scout by not reaching out to tool use and scraping every single GitHub repo and issue every single time?
No, that’s a common misconception. MoE despite the name doesn’t have an “expert” slice that’s an expert in any particular thing. It’s just trained models such that a subset of all memory weights is active on a single token. Often, the weights active on the next token are different. It’s used to improve memory bandwidth (throughput) and also to even out load in a distributed serving model - one GPU hosts one “expert” and the conversation is routed to it for the next token. This requires a lot of movement of the KV cache of course so it’s difficult to retain speed across multi node splits - usually it’s a rack of GPUs that you do this for.
MoE's in the abstract often get presented as if theres explicit layers of experts for any given domain of knowledge, like your coding tasks are being routed to coding experts, but it's really not that at all.
THe original MoE paper from Noam Shazeer et al. is worth a read on this bit, though the paper is admittedly pretty dense. But TL;DR is that each expert layer is learning highly abstract, localized structural and syntactic patterns in the data to minimize the loss function, and its doing this token-by-token (which in some cases may have some domain clustering, but that's just incidental).
When you start batching your queries, even if they all seem like theyre in a single domain, if you visualized the activations you'd notice that most if not all of the network is lighting up on the batched forward pass.
> So if I'm e.g. coding a SwiftUI app for navigation, I'd take 9B of basic coding and reasoning, add 10B of swift/swiftUI, add 5B of GIS/geography knowledge and another 5B of frontend app design knowledge.
Aren't you describing RAG or even MCP servers? Heck, nowadays you get that also with agent skills and specialized tool calling.
Definitely not MCP, as that pulls info into the context. Unless contexts become REALLY big so that I can add 10B in swift knowledge, that's not gonna help me.
Possible RAG? I don't know enough about how that works, but I think that's not quite it either. I don't want to import facts like "the swift standard library contains a reverse array function", i more want to import knowledge - e.g. the parameters used to generate the text to reverse an array in swift.
Tool calling wouldn't do it either. You'd have to encode every single possible bit of useful info into the tool call, and the tool response would have to encode every piece as well (variable names, function scopes, types defined in other files, etc). E.g. how does it find a bug, if you have to pass understanding back and forth between the brain that understands debugging and the brain that understands THIS code?
You can look at the benchmark and the GPT-5 failures like answering "April 22, 2019" instead of the correct "Oct 23, 2018" for the question:
What day, month, and year was Carrie Underwood's album "Cry Pretty" certified Gold by the RIAA?
If your idea of the smartest person in the world is the guy who always wins tuesday night pub trivia, this blog post is for you. It also gets it's foundational factual claim wrong (as seen via epoch.ai). Very on brand.
To reason properly about the human condition (eg. World War) wouldn't you need to reason on some facts ? And then reason how some "facts" change the human behaviour ? How can you arrive via pure reasoning to predict how a collective of humans act ? We are not reasonable, humans are not logical deterministic machines confined to algebraic rules.
Man, both that Blogpost and some comments here read like science-fiction.
Specifically, creative writing driven by nerds dreaming about a future, without proper grounding in reality, constraints and all that stuff.
Which is kinda ironic given the topic.
And also important to do, because we should keep dreaming. We should just also be aware of when we are doing that and mark it as such.
Reasoning is not separable from the particulars of a specific language game / linguistic practice. I’m sure some of the things the author predicts will happen, but the idea of some sort of abstractly perfect reasoner separated from the semantic content of language gets fundamentally wrong what reasoning is.
Great article, even if it will be interesting to see whether things continue to develop in such a direction or not.
> There's a version of this future where the model card stops listing a knowledge cutoff at all, because what's left in the weights goes stale on a scale of years instead of weeks.
Future?
Even just recently I’ve read of two approaches to this problem:
Cactus have come up with Needle [0][1], which is their tool-calling focused 14 MB model (still an LLM!) – no world knowledge engrained.
And instead of say, tool call structure, VibeThinker [2][3] focuses on reasoning over world knowledge.
Combine these two approaches with a reliable search tool/a safe way of accessing the internet for the model, and you’ve got a probably slightly slower model for factual questions, which on the upside however doesn’t hallucinate.
Providing not just any a baseline, but a correct and useful one, is ever more important the less the model is grounded in world knowledge – misunderstandings probably compound faster if there is no general grasp of (broadly) “life on earth”, or computers, or whatever.
And secondly, I think (consumer-oriented) search becoming worse and worse is a challenge that’s mostly solvable (but far from solved!) for the big labs: (Mostly) trusted or even editorialized/reviewed sources like published work, Wikipedia, etc. is something they could index internally, it doesn’t need to come from a random blog site on the public internet. Furthermore, there’s a whole slew of companies specializing in crawling-for-LLM (i.e., bypassing bot protections) now as well.
the technicals of search did not get worse. the economics of the search provider did. If government wasn't stuck in the 50s mindset, search provider would have been nationalized just like a utility.
Current AI is like the film company producing TV series or movies
Your question is like a story outline. You tell the film company that this is the movie you want. The AI film company then searches for existing similar stories. If similar stories do not exist or details are missing, screenwriters use imagination to fill in the gaps (remember hallucination? It's just a makeup.)
Indeed !
LLM are creative writers, not journalists.
Relying on overfitting for factual accuracy in not tenable.
I don't understand why grounded RAG with judges in not the norm.
The author is factually incorrect here. Moving information out of the model weights and into the input of the model's context window in no way ensures that the model will accurately output content that was input from the context. This is true even when RAG is used to input exactly the correct data.
LLMs work is being intelligent not having knowledge of everything is ok. But, they have to be intelligent enough (with some degree of knowledge) that where to find the information (search tools or any other tools for that matter)
Not too long ago, the claim was that having multiple Phds worth of knowledge was the real value of AI, but now that's gone and it's suddenly a good thing.
With all the focus on coding and agentic use, I wonder if the rest of the world will notice or care? Most AI use is not for coding or harnesses contrary to what this website thinks. Maybe this is how American companies stay winning.
For example I prefer Kimi K2.6 1T parameter to Flash V4 0731 230B parameter, even if it is less intelligent.
What is the rest of the world using LLMs for? Agentic stuff seems pretty universal. Obviously coding stuff is only interesting to people who want to code, but automating complex digital tasks seems useful in all kinds of contexts.
I’m working at a company where everyone is using LLMs for everything and I’m not aware of anybody using anything truly genetic. It’s all human in the loop babysitting the tool.
I agree that most AI use in terms of users may not be for coding or agentic use (everyday people are asking chatgpt for something or looking at google ai summary), but with respect to AI usage, I speculate that the vast majority of usage is coding and agentic because they're super token hungry.
In terms of the value proposition of AI replacing knowledge workers, all value is in coding agents (coding agents as general agents).
This makes hallucination detection more important.
There's no reason that an LLM should have a vast number of obscure facts encoded. It can go out to a search engine for such facts. But the LLM has to be clear on what it doesn't know.
(Google's pricing for search from programs starts at $2.50 per 1,000 queries. If an LLM reaches out to Google, it has to pay.)
> Ask the same models a plain factual question and the picture flips. On SimpleQA, a benchmark of factual recall with no tools allowed, the current leader is Gemini 2.5 Pro at 53%, so the best recall money can buy still misses half the questions.
Maybe I'm missing something, but on the page of SimpleQA there are only 34 models, none from Anthropic, the most recent one from OpenAI is GPT 4.1 (or O3 mini?), no Deepseek V4, no GLM.. So it turns out that in a choice of small or outdated models, Gemini 2.5 flash shines. Hardly a starting point for an argument.
> the 24GB card that's been sitting in gaming PCs since 2022.
I'd wager most people have less. In 2022 a 3080 might have 12 GB if you were lucky, 10 if you weren't -- and you paid for the privilege. A current RTX 5080 is only 16GB.
I hope there is interesting research being done to stop models from 9B to 3T from authoritatively giving the wrong answers to factual questions instead of doing tool calls, but this article will tell you nothing about it. That's because it's clearly prompted by someone who simply noticed this deficiency very obvious to all practitioners and even users.
In my experience, it feels like this problem has been mostly solved. When I interact with a chatbot, I see it often uses web search to answer my questions and hallucinates much less then they used to. Furthermore, I suspect a lot of gains are merely from the system prompt instructing the chatbot to use web search to verify all facts it says.
Ideally what I'd like to see is pluggable knowledge bases.
So if I'm e.g. coding a SwiftUI app for navigation, I'd take 9B of basic coding and reasoning, add 10B of swift/swiftUI, add 5B of GIS/geography knowledge and another 5B of frontend app design knowledge. My model doesn't need to know a single line of python.
Then when I want to research electronics components, I grab a 15B model of agentic research techniques, and add in 10B of electronics knowledge, etc.
I don't want general purpose models. They try to be everything to everyone. I want to click together a model that is laser-focused on what I am doing, and I want to run it locally
An LLM works better the more disparate world knowledge it has, even if it's not immediately obvious why it would be relevant. The model finds a structure to the problem you give it in a largely language-agnostic way that benefits from training on every language (these things are direct descendants of Google Translate), and even non-programming knowledge - the structure of your task might resemble an ancient Chinese poem that influences the model's response, for example. That structure is considered a form of compression, as some fascinating and illuminating recent 3blue1brown videos get into - a common pattern in Haskell or FORTRAN and a situation described in an ancient Chinese poem may all compress to something quite similar to your task, thus when the model compresses the idea of your task it immediately draws from those ideas.
There are "experts" which do divide parts of the model that are found to activate together for specific tasks, so they can be processed in parallel to join the result at the end, but it's nowhere near the granularity of a SwiftUI expert and a python expert. The difference in those things is so trivial from an abstract point of view that it would make no sense. They would be 99% the same.
Distillations also come into this but I'm highly skeptical you could make one guaranteed to only know programming and only in one programming language (especially with as small a sample set as SwiftUI relative to something like C) without its efficacy being hobbled by tunnel vision. Reminiscent of the SpongeBob episode where he empties his mind of everything except fine dining and breathing, then can't remember his name and goes insane. Beyond the basic concepts of general coding and the trivia of syntax, getting anything done requires a large intersection of disparate world knowledge and the ability to apply it to new situations.
This makes a ton of sense, and seems like it's coming from somebody who understands LLMs better than I do so I will definitely take your word for it - but it doesn't totally track with my experience of running small local models either. In my experience some models are better at language X and others better at language Y - and all of them are better at language X and Y than language Z-I-just-made-up. I take that to mean there is some relationship between training data and skill. Maybe not enough to make pluggable modules, tho...
Thanks for putting this so well. The mathematical evidence for this “general intelligence underlying everything” is the “universal geometry of embeddings” paper. Fascinating read. Or as the ancient philosophers used to say, the one who knows God knows everything.
This is a fundamental misunderstanding of how LLMs work. You can’t really specialize a model. You specialize the harness. A well-trained general purpose LLM doesn’t need examples in its training data, it can write good code in a new language you invented yesterday with just a spec definition. And it will perform better than a small model trained on lots of examples of your invented language. The reason is because of the “universal geometry of embeddings”, i.e all human languages have the same underlying pattern structure, so any model that is very good in any language is good in all languages. Attempting to specialize a model for a particular purpose often decreases overall performance. Fine-tuning is just a hack to make dumb models more reliable on limited tasks but they become incapable of doing anything else. Unless you are building a factory assembly line where a model is literally doing the same thing over and over, you almost always want a general purpose model over a specialized one.
This is so right. We training Whisper Large model on 20,000 audio samples specific to a domain and it ended up reducing the ASR by 5% while improving WER of the finetuned domain by 0.5%.
Instead we ended up with no finetuning. We give audio snippet to 2 AsR models, take 3 best transcriptions and ask the LLm to pick the best based on the context. That produced significantly higher accuracy in how an agent understands the users.
> My model doesn't need to know a single line of python.
If I had to guess, the weights necessary to encode "how to program" are much larger than the final step of "output python."
<THAT'S A BINGO>
ie what everyone asking for this fails to immediately realize.
But the understanding of the language library ecosystem, or even better, your codebase, could let it execute faster and with less context usage.
The problem with this idea is that knowing Python makes the model a better Swift programmer, as does a higher-number of parameters during training. So you'd be so much better off with a 90B general purpose model trained on everything anyway.
> I want to click together a model that is laser-focused on what I am doing
This is roughly what multi-agent systems are built for.
This is possible with models too, but "making one on the fly" is much easier with agent coordination rather than model weights, since they all speak the same language.
There is an IBM Mainframe vs Google Distributed system division here. Like Seymour Cray said - two oxen or 1024 chickens.
Chickens are harder to harness, so a lot of my work is in sled-dog territory for agent harnesses & command structures.
> This is roughly what multi-agent systems are built for.
I think I disagree. For some things, maybe that works - but think of a multi-agent system where one agent understands the code, and passes it off to the reasoning agent to figure out what the bug is. This system is going to suck. Because encoding enough info to figure out what the bug is would just be dumping every single line of the code.
So say agent 1 (reasoning) asks agent 2 (swift) to explain what is happening in File.swift. Anything agent 2 passes to agent 1 short of the entire code is a lossy transfer - and then the bug gets missed.
Sounds like unix philosophy. Or like Neo downloading Kung-Fu. Good either way.
everything is a token
There is no token
If that is better (and possible) it will be baked in the tools. But is a model that doesn't know python better or worse at swift is what I wonder.
Sounds like MoE but more pluggable. Not sure if anyone is researching something like this. I still think your swift or GIS model will need basic reasoning and coding to work, so it's more like multiple smaller models which you can load as needed, e.g. sub-agents for GIS needs the GIS model.
Another approach would be to have basic coding and reasoning model and then load specification for language and libraries into context, it could work for self-hosted models, but I don't want whole specification of the language to be send to API and waste tokens on that.
My motivation is somehow the model should be able to work without hitting an external resource every single time and somehow have all the knowledge necessary locally so we don't get rate limited but how can you stay up to date with the latest information while still remaining a good boy scout by not reaching out to tool use and scraping every single GitHub repo and issue every single time?
This would defeat the AGI narrative/belief that so many building these models have
Would it? AGI was always about the efficiency gains of versatility, not strictly all forms of efficiency
AGI is whateve rmakes the VC open their wallet.
On the otherhand, your own brain probably doesn't use your eyes to hear.
You don't think reading lips helps you understand people?
That's still not hearing.
Hearing has volume, direction, pitch, it's spacial processing etc
You could if you do lip reading
Admittedly I’m pretty ignorant of the details, but I thought this was the mixture of experts architecture
No, that’s a common misconception. MoE despite the name doesn’t have an “expert” slice that’s an expert in any particular thing. It’s just trained models such that a subset of all memory weights is active on a single token. Often, the weights active on the next token are different. It’s used to improve memory bandwidth (throughput) and also to even out load in a distributed serving model - one GPU hosts one “expert” and the conversation is routed to it for the next token. This requires a lot of movement of the KV cache of course so it’s difficult to retain speed across multi node splits - usually it’s a rack of GPUs that you do this for.
MoE's in the abstract often get presented as if theres explicit layers of experts for any given domain of knowledge, like your coding tasks are being routed to coding experts, but it's really not that at all.
THe original MoE paper from Noam Shazeer et al. is worth a read on this bit, though the paper is admittedly pretty dense. But TL;DR is that each expert layer is learning highly abstract, localized structural and syntactic patterns in the data to minimize the loss function, and its doing this token-by-token (which in some cases may have some domain clustering, but that's just incidental).
When you start batching your queries, even if they all seem like theyre in a single domain, if you visualized the activations you'd notice that most if not all of the network is lighting up on the batched forward pass.
I would love this but I think the General Reasoning and Make No Mistakes modules would be massive.
I mean, I get pretty damn good general reasoning out of Qwen3.8-27B quantized to 4bits, and that knows Swift, Python, Node, Ruby, Rust, etc. etc. etc.
And so far even the biggest model doesn't seem to have a working Make No Mistakes module, so maybe that's not needed
This is going to be how I write my resume now: how many billion parameters of knowledge I have on each topic
This is possible today with an agent such as OpenClaw and hermes
I strongly suspect this will be the future
> So if I'm e.g. coding a SwiftUI app for navigation, I'd take 9B of basic coding and reasoning, add 10B of swift/swiftUI, add 5B of GIS/geography knowledge and another 5B of frontend app design knowledge.
Aren't you describing RAG or even MCP servers? Heck, nowadays you get that also with agent skills and specialized tool calling.
I don’t think so because those both live in the context window and as such pollute it when they’re not performing optimally.
I think having unused or rarely used weights doesn’t influence the results as poorly as RAG injecting irrelevant facts.
It sounds to me like some sort of “dynamic MoE” where you can add/create or remove experts on the fly.
I think what you’re describing is the closest approximation we reasonably have right now though.
I don't think so?
Definitely not MCP, as that pulls info into the context. Unless contexts become REALLY big so that I can add 10B in swift knowledge, that's not gonna help me.
Possible RAG? I don't know enough about how that works, but I think that's not quite it either. I don't want to import facts like "the swift standard library contains a reverse array function", i more want to import knowledge - e.g. the parameters used to generate the text to reverse an array in swift.
Tool calling wouldn't do it either. You'd have to encode every single possible bit of useful info into the tool call, and the tool response would have to encode every piece as well (variable names, function scopes, types defined in other files, etc). E.g. how does it find a bug, if you have to pass understanding back and forth between the brain that understands debugging and the brain that understands THIS code?
You can look at the benchmark and the GPT-5 failures like answering "April 22, 2019" instead of the correct "Oct 23, 2018" for the question:
What day, month, and year was Carrie Underwood's album "Cry Pretty" certified Gold by the RIAA?
If your idea of the smartest person in the world is the guy who always wins tuesday night pub trivia, this blog post is for you. It also gets it's foundational factual claim wrong (as seen via epoch.ai). Very on brand.
https://epoch.ai/benchmarks/simple-qa-verified?view=graph&ta...
https://logs.epoch.ai/inspect-viewer/c79c08da/viewer.html?lo...
But is Reasoning and Facts truly separate ?
To reason properly about the human condition (eg. World War) wouldn't you need to reason on some facts ? And then reason how some "facts" change the human behaviour ? How can you arrive via pure reasoning to predict how a collective of humans act ? We are not reasonable, humans are not logical deterministic machines confined to algebraic rules.
Man, both that Blogpost and some comments here read like science-fiction.
Specifically, creative writing driven by nerds dreaming about a future, without proper grounding in reality, constraints and all that stuff.
Which is kinda ironic given the topic. And also important to do, because we should keep dreaming. We should just also be aware of when we are doing that and mark it as such.
Reasoning is not separable from the particulars of a specific language game / linguistic practice. I’m sure some of the things the author predicts will happen, but the idea of some sort of abstractly perfect reasoner separated from the semantic content of language gets fundamentally wrong what reasoning is.
Great article, even if it will be interesting to see whether things continue to develop in such a direction or not.
> There's a version of this future where the model card stops listing a knowledge cutoff at all, because what's left in the weights goes stale on a scale of years instead of weeks.
Future?
Even just recently I’ve read of two approaches to this problem:
Cactus have come up with Needle [0][1], which is their tool-calling focused 14 MB model (still an LLM!) – no world knowledge engrained.
And instead of say, tool call structure, VibeThinker [2][3] focuses on reasoning over world knowledge.
Combine these two approaches with a reliable search tool/a safe way of accessing the internet for the model, and you’ve got a probably slightly slower model for factual questions, which on the upside however doesn’t hallucinate.
[0] https://cactuscompute.com/needle
[1] https://news.ycombinator.com/item?id=49246804
[2] https://arxiv.org/abs/2606.16140
[3] https://news.ycombinator.com/item?id=48639240
That kind of setup is super dependent on a search engine, and search keeps getting worse.
Generally I agree, with two caveats/thoughts:
Providing not just any a baseline, but a correct and useful one, is ever more important the less the model is grounded in world knowledge – misunderstandings probably compound faster if there is no general grasp of (broadly) “life on earth”, or computers, or whatever.
And secondly, I think (consumer-oriented) search becoming worse and worse is a challenge that’s mostly solvable (but far from solved!) for the big labs: (Mostly) trusted or even editorialized/reviewed sources like published work, Wikipedia, etc. is something they could index internally, it doesn’t need to come from a random blog site on the public internet. Furthermore, there’s a whole slew of companies specializing in crawling-for-LLM (i.e., bypassing bot protections) now as well.
the technicals of search did not get worse. the economics of the search provider did. If government wasn't stuck in the 50s mindset, search provider would have been nationalized just like a utility.
I think the idea is reasonable, however the SimpleQA Bench stopped measuring in Sept. 2025.
So newer data would be interesting.
(It seems a bit like an AI generated argument that uses old facts - something that happens to me quite often)
There’s SimpleQA Verified
https://epoch.ai/benchmarks/simple-qa-verified
> This mostly solves hallucination
Current AI is like the film company producing TV series or movies
Your question is like a story outline. You tell the film company that this is the movie you want. The AI film company then searches for existing similar stories. If similar stories do not exist or details are missing, screenwriters use imagination to fill in the gaps (remember hallucination? It's just a makeup.)
So you cannot solve hallucination of AI
Indeed ! LLM are creative writers, not journalists. Relying on overfitting for factual accuracy in not tenable. I don't understand why grounded RAG with judges in not the norm.
A sharp critical thinker draws admiration, but it is a joy to talk to some dude who knows stuff.
> This mostly solves hallucination
The author is factually incorrect here. Moving information out of the model weights and into the input of the model's context window in no way ensures that the model will accurately output content that was input from the context. This is true even when RAG is used to input exactly the correct data.
Intelligence vs Knowledge
LLMs work is being intelligent not having knowledge of everything is ok. But, they have to be intelligent enough (with some degree of knowledge) that where to find the information (search tools or any other tools for that matter)
Not too long ago, the claim was that having multiple Phds worth of knowledge was the real value of AI, but now that's gone and it's suddenly a good thing.
That's a sales tactic -- not a logical position.
With all the focus on coding and agentic use, I wonder if the rest of the world will notice or care? Most AI use is not for coding or harnesses contrary to what this website thinks. Maybe this is how American companies stay winning.
For example I prefer Kimi K2.6 1T parameter to Flash V4 0731 230B parameter, even if it is less intelligent.
What is the rest of the world using LLMs for? Agentic stuff seems pretty universal. Obviously coding stuff is only interesting to people who want to code, but automating complex digital tasks seems useful in all kinds of contexts.
I’m working at a company where everyone is using LLMs for everything and I’m not aware of anybody using anything truly genetic. It’s all human in the loop babysitting the tool.
What tool?
Even basic clients are now harnesses. A lot of chat interfaces are using memory systems, web search and other stuff under the hood.
Not as agentic as openclaw, but not a straight closed conversation either.
Learning new things for example, general curiosities, text processing
I agree that most AI use in terms of users may not be for coding or agentic use (everyday people are asking chatgpt for something or looking at google ai summary), but with respect to AI usage, I speculate that the vast majority of usage is coding and agentic because they're super token hungry.
In terms of the value proposition of AI replacing knowledge workers, all value is in coding agents (coding agents as general agents).
This makes hallucination detection more important.
There's no reason that an LLM should have a vast number of obscure facts encoded. It can go out to a search engine for such facts. But the LLM has to be clear on what it doesn't know.
(Google's pricing for search from programs starts at $2.50 per 1,000 queries. If an LLM reaches out to Google, it has to pay.)
I think the biggest problem with the models is they don’t actually have any decent lookups except chunked document embedding search
> Ask the same models a plain factual question and the picture flips. On SimpleQA, a benchmark of factual recall with no tools allowed, the current leader is Gemini 2.5 Pro at 53%, so the best recall money can buy still misses half the questions.
Maybe I'm missing something, but on the page of SimpleQA there are only 34 models, none from Anthropic, the most recent one from OpenAI is GPT 4.1 (or O3 mini?), no Deepseek V4, no GLM.. So it turns out that in a choice of small or outdated models, Gemini 2.5 flash shines. Hardly a starting point for an argument.
This article is AI output.
> the 24GB card that's been sitting in gaming PCs since 2022.
I'd wager most people have less. In 2022 a 3080 might have 12 GB if you were lucky, 10 if you weren't -- and you paid for the privilege. A current RTX 5080 is only 16GB.
I'm still getting by mostly fine with a 4GB 1650 Super. (2020 vintage.)
I hope there is interesting research being done to stop models from 9B to 3T from authoritatively giving the wrong answers to factual questions instead of doing tool calls, but this article will tell you nothing about it. That's because it's clearly prompted by someone who simply noticed this deficiency very obvious to all practitioners and even users.
In my experience, it feels like this problem has been mostly solved. When I interact with a chatbot, I see it often uses web search to answer my questions and hallucinates much less then they used to. Furthermore, I suspect a lot of gains are merely from the system prompt instructing the chatbot to use web search to verify all facts it says.
> Models Are Getting Dumber on Purpose
I know is editorialized, but a more accurate title to this content would be either :
Models Are Getting Ignorant on Purpose
or
Models Are Getting Less Knowledgeable on Purpose