This isn't as easy as it sounds. Every ML model is struggling to balance between generalization and test performance.
Taking a good model like GLM5.2 and just fine tuning it on coding can decrease real world performance due to mechanics like catastrophic forgetting. There is also other interesting behaviors were training on a broad training set can improve coding performance because there is positive transfer.
There is 100% an effort to make solid coding focused models, but it is very hard to do that without including capabilities across a broad set of adjacent tasks.
Help me understand this viewpoint that AGI being possible in the near-ish future is a myth, I see it repeated quite a lot.
I've been in NLP since the LSTM days and it's hard for me to look at LLMs and not just think they are incredible. It's truly a different level of expressiveness. So much of capabilities research is pointing to LLMs effectively learning a world model.
RLVR is also proving really effective. It is hard for me to imagine a world in the future where LLMs aren't at human level performance across a wide variety of tasks.
I fully acknowledge that current LLM labs have a financial interest in people believing AGI is very near, but from what I'm reading in the literature and seeing myself experimenting with the SOTA models it doesn't seem totally unreasonable.
What evidence are you seeing that makes you confident that AGI in the soon-ish future is a complete myth?
There doesn't really seem to be anything of substance in the actual executive order.
Section 1 doesn't say anything
Section 2 seems to boil down to: "improve cyber security and maybe use AI if we can find funding for it"
Section 3 proposes building a benchmark for evaluating cyber security performance of models that developers can choose to benchmark against. This seems like a good idea, I know Jack Clark has been a huge advocate for government's getting in with benchmarking.
Section 4 says to prioritize prosecuting cyber crimes. Not sure why they wouldn't already be prosecuted.
Are we looking at the same data? On that site I see that opus 4.7's and gpt 5.5's g scores are within each others confidence intervals, and both significantly ahead of the number 3 model.
Your comment makes it sound like they are miles apart, which the benchmark doesn't seem to support.
Edit:
I looked at the data more and the two models are only basically equal when looking at the mean of all the tests. Gpt 5.5 significantly outperforms opus 4.7 in coding, while opus 4.7 significantly outperforms in "decision making." I'm not seeing details on what decision making explicitly means.
I had a similar experience recently, where I logged in to Facebook after not using it for years and was shocked by how much garbage was there. My spouse does use Facebook somewhat regularly so I looked at her feed and it was much more reasonable.
I wonder if for those of us that haven't used Facebook in years the recommendation algorithm is essentially default. Which much like the default youtube algorithm, is completely garbage. But if we did use it (which I have no intention of doing), it would start being more reasonable.
I worked on something very similar for my master's degree.
The problem I could never solve was the speed, and from reading the paper it doesn't seem like they managed to solve that either.
In the end, for my work, and I expect for this work, it is only usable for pre generated terrains and in that case you are up against very mature ecosystems with a lot of tooling to manipulate and control terrain generation.
It'll be interesting to see of the authors follow up this paper with research into even stronger ability to condition and control terrain outputs.
There are still some features that a miss from Google photos. There isn't any way (that I know of) to auto add pictures to an album based on the face. I used to have dedicated albums for family members, and it was nice to have the auto updated.
Face recognition in general just isn't as good as Google Photos.
It's still an amazing piece of software and I'd never go back, but it isn't perfect yet.
Diffusion LMs do seem to be able to get more out of the same data. In a world where we are already training transformer based LLMs on all text available, diffusion LMs ability to continue learning on a fixed set of data may be able to outperform transformers
This exactly. For parents it is not a choice, you absolutely must have a parent sitting by a young child. The effect of not automatically putting parent and children next to each other would just be making tickets more expensive for parents.
Taking a good model like GLM5.2 and just fine tuning it on coding can decrease real world performance due to mechanics like catastrophic forgetting. There is also other interesting behaviors were training on a broad training set can improve coding performance because there is positive transfer.
There is 100% an effort to make solid coding focused models, but it is very hard to do that without including capabilities across a broad set of adjacent tasks.