bluejay2387

This might be one of the most surprising things I have seen in AI. It's like we taught someone to make vases by reading an enormous library of "How to make a vase" books and somehow that person intuited material science.

Is this just giving a general-purpose agent a workspace full of documents and an assignment, then seeing what happens? I feel this doesn't really say anything. It appears they mostly dumped all of this in base models with little to nothing in the way of infrastructure to support it. If you just randomly grab some person off the street and say 'do taxes' they would probably mess up just as badly. The interesting question would be what the models could do in a well engineered harness focused on taxes. That would be reflective of the performance boundary and how serious we would have to take AI in that field.

There is a natural limit to the size of a corporation (it's big) before the inefficiencies of size start to choke out your ability to get anything done, and AI doesn't fully fix that. You can't do an arbitrarily large number of things and expect to keep running as a business, this is why even large firms today tend to break into separate companies (a lot of the large companies you can think of actually run as multiple smaller companies under one umbrella). Even when AI is 'as smart as us' it's not going to remove this limitation (might lessen it)... see irreducible complexity, diseconomies of scale, etc...

The claims seem to be an overreach of what is actually proven. We've known about progressive feature complexity and the impacts of early training structures on later training for a really long time (since CNN's, The Lottery Ticket Hypothesis, etc...) and neither of those phenomenon are mutually exclusive with parameter overloading. The Multimodal Gap work (https://arxiv.org/abs/2203.02053) seems to imply networks have a tendency towards condensed representations that are clearly not semantic and a bunch of other considerations are already on the table for discussion. I don't think these have all been considered sufficiently to start claiming we are around the corner from mechanistic interpretation.

Anyone else feel ripped off? There are solar systems with 6 possible Earth like planets and others with Super Earth Garden Worlds and what do we get? One measly mediocre Earth with our nearest neighbors being a poisonous rock and boiling ball of acid.

If I took Metasploit and hooked it to a Cron job then used it to hack HuggingFace they would lock me away for life and the Justice Department wouldn't care that I use software to automate it. No idea why these companies get away with what is essentially the same thing.

But I can probably prompt Claude better than your customers too, so can at least 50 of the people I know (advantage of a tech heavy social circle). Good luck turning your prompt into a business when me and 50 of my friends can replicate your business model in about 10 second. Don't believe? Go ask Jev devs how they feel about Clef.

"I have in mind something like a “deployable intellectual reserve”: communities of mathematically sophisticated people that humanity can call upon to help understand consequential AI-enabled breakthroughs. " -- I can't be the only one who thought about the scene where the scientists ask Deep Thought the ultimate question...

Can we all agree to ban the term 'non-autoregressive model' from our collective vocabulary? It's like naming a database a 'non-spreadsheet'. Decision models, direct inferenceing models, etc... there are sooo many better names for this. I would have thought we had learned our lesson with the whole 'Vectorless RAG' incident. We should go back and check what they named the Gorilla models at Berkley.

It took us a few hours to implement that one we used in 2022. This isn't about usability its about a huge population working on this stuff not really knowing what is available until it becomes a meme.

"I don't see much substance to this buzz..."

Agreed. This isn't new. I led a research team at a Fortune 500 that used a transformer based classifier approach in a commercial product as far back as 2022 and we didn't come up with it. It was already common enough that we found the inspiration for our implementation on some web forum. Models like RouteLLM have been around for a long time. The news here isn't that a new model type came about, its that a large percentage of people messing around with this stuff that are new to AI just learned that not all transformer based implementations need to be autoregressive.

I really like this game and have played it for years but I think it's popularity outgrew it's niche. It's a very niche game, the devs have made the ultimate space Lego kit, but I see lots of people complain about it being boring because there are no stakes, no challenge. I think they are trying to play the game they want it to be not the game it is. The overwhelming scale of the meme this game has become I think is the cause of that.

I don't see not having AI as a feature, but I'd prefer for them to take the approach of optional, easy to implement interfaces that increase the ability for the suite to be 'scripted' rather than implement AI directly. My motivation is that the field of AI is moving so fast I think its a better long term plan to make your software something that can be automated with AI than to try to fight the speed of that innovation curve. I get that even that can be difficult because of security concerns, but it seems more straightforward than implementing models directly into your application.

I am sure it will be fantastic for the whole 6 seconds before it blows my weekly usage cap.

In this case, the fully open source pipeline is probably as valuable or more so than the weights, so releasing early has some justification.

I know I am in the overwhelming minority and that most people will just continue to buy Sony's stuff anyway, but I decided a few years ago when all the ownership crap came up that I am just not buying stuff that I don't own. Sony's PS4 went to my local electronics recycler and I haven't bought a Sony product since. I can't realistically do it with everything, there are some products I have to buy rather I get to own them or not. But I am not putting up with it where I can and gaming is optional -- plus GOG has served me really well since.

I've worked with a number of clients that have lost senior engineers due to old age or retirement, and the impact is often very significant regardless of how sophisticated their knowledge management infrastructure is. Documentation can not fully replace the insight and situational awareness many of these experts posses and most of the time knowledge management still suffers from the 'not knowing what you don't know' issue. Only well structured 'apprenticeship' programs come close. I know some are going to point to AI, but I've worked in AI for over 2 decades and have extensive experience in Expert Systems and Knowledge Engineering from early in my career. It is not a complete solution. Given today's technology its trivial to replicate 'text book knowledge' but the difficult, expensive problems occur in the margins and its still hard to train AI on open ended / rare occurrences outside of domains that have specific advantages like software engineering. Talented and experienced people are still required. Maybe that will change soon, but AI isn't there yet.

I am guessing the developer took into account the fact that the younger generations these days don't have the patience for real time defragging.

I would imagine there is some obvious bias and that the AI backlash plays a role, in that many of those that would be buying the models are likely to be part of industries where we are seeing the most backlash against the technology. In another train of thought, how mind blowing is it that 4 years ago I could barely get 512px images out of models and just 4 years later there are legitimate arguments as to if AI generated 3D models are usable for production work?

I think the attacks generated by Meta, Open AI and Anthropic prove that large corporations are not responsible enough to be trusted with advanced AI, so we should ban all commercial AI services and only allow open source models that are in the hands of hobbyists and individuals -- hobbyists and individuals that have so far proven to be much more trust worthy.

I agree, the whole 'developers don't code' AI defense was pretty ridiculous. I also concur that development is going to get a lot harder. Anyone that has successfully used AI coding tools knows that you can get massive productivity increases but now I have to figure out how to get a bunch of hyper active child like coding entities with memory deficiencies to build stuff without going off the rail and introducing massive security problems or building something completely mismatched to the requirements. If you can do it, yeah you can get 10x results. But now I am engineering harnesses, architecture specifications, agent structures, statistically sampling, and formal verification systems to guide the code instead of writing the code.

"For decades, logic and CS researchers have known what reasoning is." ... this is a fairly significant overstatement. There is not complete agreement on this term and our understanding continues to evolve. The models don't have to think like humans to think.

Saying that LLM's only offer an 'approximation' of reasoning is also an overstatement as it is not a resolved topic.

But to the original point, its not exactly just semantics if thought traces are not doing the job that they were originally thought to do. There is value in knowing how these things actually work. If chain of thought is just grounding the latent space and not directly contributing to the process of generating a response it has implications on how we test and verify the reliability of models if nothing else... doesn't mean they aren't useful but it definitely impacts many of the tools we could have to evaluate their performance.

I think this should probably be scoped to 'generic router systems that don't understand query context' are not useful. We have had lots of good results with routers that understand the context of the types of workloads they process and can route requests to the most efficient models.

Better than Cities: Skylines II turned out to be.

I think a lot of were already using Mermaid and/or Python/Matplotlib(etc) for this. What would be the advantages to Flint?

"End times are nothing new, it's the historic default mode." -- might be the smartest thing I have read in many weeks.

The entire domain of NP-Complete problems would beg to differ with you.

About 90% of my coding is on Qwen 3.6 27b and Open Code with some custom skills and Semble. It is NOT as smart as CC or Codex but its enough to get most of my work done. I didn't set out to replace CC and Codex (I have an RTX 6000 so the TPS is faster than I care about, but the RTX 6000 was originally for other work). I only tried this just to see how close you could get to a frontier model for coding as an experiment, but it was good enough that I stuck with it. I still fall back to Codex for really complicated stuff and to polish UI's as that seems to be the weakest element to working in Qwen.This isn't a recommendation because I don't think most people have an RTX 6000 laying around and the cost would be many years of MAX CC or Codex subscriptions, but at least this seems possible. Maybe in a few more years it will even be practical.

Other Notes: I have had to set the compact target to 75% on a 256k context window as once the conversation length goes about 100k I start seeing a drop in the quality and speed. This becomes very problematic after about 150k. I tried Qwen 3.5 122b too but it actually seems much worse at coding than 3.6 27b even though its much larger. Maybe because I am using a 4bit quant or maybe I just don't have it configured correctly? I know 3.6 is newer but I didn't expect it to out perform a model that is much larger from the prior generation. Gemma 4 31b is a good model for other tasks but at least my personal experience is that Qwen outperforms in coding. Nemotron Super 120b is great at a lot of stuff but it also seems to be not as good at coding as Qwen. This was very surprising to me.

In the US -- once our nation finishes attacking our own education system -- this is definitely something a group of academic institutions could get together and accomplish. I assume the same is true in other countries. Companies like Nvidia and AMD might even support that effort, as they make money on the hardware and would probably be more than happy for there to be more reasons to use it. There may have not been a compelling enough motivation to achieve this before, but "models" didn't have this level of strategic relevance until relatively recently. Nvidia has been fairly good about releasing open weight models in the last few months.

From what I can tell the majority of developers here have moved into the "Anger" stage.