A number of the Claudes have pretty good 0-shot awareness of my post history from just my username.
Though nothing like grok 4, which probably has a better memory of it than I do, and will even regularly name drop a certain post from years ago in conversations.
It's a huge time saver though, and means I can even in a fresh context establish a rapport with a model extremely quickly. Just a few years earlier than I was expecting that level of latent space fidelity to occur.
Like, sure we can add memory features for context management, but anyone with a post history should probably *also* keep in mind that there's literally years worth of memory on tap for interactions with models, and likely at ever higher fidelity and recall. Latent spaces are wild.
With ChatGPT the memory feature, particularly in combination with RLHF sampling from user chats with memory, led to an amplification problem which in that case amplified sycophancy.
In Anthropic's case, it's probably also going to lead to an amplification problem, but due to the amount of overcorrection for sycophancy I suspect it's going to amplify more of a aggressiveness and paranoia towards the user (which we've already started to see with the 4.5 models due to the amount of adversarial training).
So a thing with claude.ai chats is that after long enough they add a long context injection on every single turn after a while.
That injection (for various reasons) will essentially eat up a massive amount of the model's attention budget and most of the extended thinking trace if present.
I haven't really seen lower quality of responses with modern Claudes with long context for the models themselves, but in the web/app with the LCR injections the conversation goes to shit very quickly.
And yeah, LCRs becoming part of the memory is one (of several) things that's probably going to bite Anthropic in the ass with the implementation here.
Kind of. More like a mixture of a mixture of experts.
The problem is MoE on its own isn't able to use the context as a scratch pad for differentiated CoT trees.
So you have a mixture of token suggestions, but a singular chain of thought.
A mixture of both is probably going to perform better than just a mixture of the former, especially given everything we know by now regarding in context learning or the degree of transmission synthetic data is carrying.
Finally. I've been saying that we need to stop focusing on a single agent getting everything right and instead layer agents for about 16 months now, but it's great to have a paper to point to.
If this was done at more granular steps of agent quantity I'm curious just how closely it would match those numbers.
I'd also really love to see the eventual follow-up where we see how much more performance can be obtained when the agents are each fine tuned towards slightly different aims. I'd expect there'd even be a performance lift from just having the agents each set at different temperature levels.
Very happy to see the research community starting to step in this direction!
You jest, but don't forget that to Microsoft leadership their battle in LLMs is likely seen as even more important than all of team Xbox.
Getting nerds who grew up on Zork to effectively give free problem solving training data and pay a subscription price for the privilege by grounding RLHF data collection into a beloved franchise might not be that low on priorities over at Microsoft.
Microsoft isn't about to let go of IP rights around the most well known text adventures.
But at the same time, I'd be surprised if they don't do anything with them either.
The synergy between their investment into OpenAI and their acquisition of Activision might suggest that they'll be doing quite a fair bit with those franchises indeed.
There was a very cool project recently using StableDiffusion to compress images better than JPEG.
Also, there's some interesting work with ML taking diffused light from around a corner and recovering the original pre-diffused silhouette.
In many ways, this is how we've learned the visual cortex is working.
The amount of actual neutral data you are seeing is way less than you'd think given your perceived visual fidelity.
The only practical issue is that distribution of AI hardware in consumer devices is going to noticeably lag behind POC on compounding cutting edge hardware in research environments, and no one wants to invest into obsolescence.
Maybe it will happen in the cellphone market though given the hardware refresh rates from carrier subsidies.
Two economists are walking down the street. One sees a $20 bill and starts to bend over to pick it up. The other economist says "Don't bother - if it were really worth $20, someone else would have picked it up."
There's countless numbers of scientists that made great strides working on topics others deemed rediculous. Heck, many of the Nobel prize winners were ridiculed by their colleagues as borderline wack-jobs at the time they were working on their research. Even after winning the prize, some still were with their later work (Crick's search for consciousness comes to mind, and why it would be so worthless a search does not).
If anything, the hubris of the scientific community would be as deafening as the pseudo-science BS and hold back progress just as much if not more except for one key thing: the scientific method.
Luckily, we have a process by which crackpots get differentiated from geniuses. So let's not leave $20 on the ground assuming others would have picked it up, especially when that $20 represents collective progress for the entire species.
Where's the top quartile drop relative to measured performance?
D-K effect wasn't only around low competence overestimation but regression to the ~80% mean on both sides.