I think the board (and disappearing letters) are made with a list of words in mind. Thus, to complete the puzzle (dissolve all letters) you cannot count the bonus words as they were not accounted when creating the letter graph.
Its unclear to me why the color space is 2-dimensional. Why wouldn't it be a 3-dimensional space, indexed by how much each of the 3-cones is activated ? Not clear to me from the article!
Well, the most common example si precisely screens, no? A screen displaying the color yellow is actually a spectrum of red and green peaks, stimulating your red and green cones just like a spectrum containing a single frequency of the color yellow.
String theory is in fact falsifiable contrary to popular belief. It's just not practically falsifiable with current (and likely future, for a while) technology as the energy scales we need to probe to falsify it are astronomically large.
I think the capex being fixed assumes you can just stop training the next model. But its not clear that you can afford to do that and keep selling tokens.
And if capabilities plateau such that training the next one is useless, then the margins will drop fast due to competition.
If you combine the LLM probability distribution with arithmetic coding you can actually use them to compress text losslessly. When people reports 'bits per byte', it is actually the compression rate for text.
GPT-2 for instance achieves roughly 1 bit per byte, so it can be used to compress (english) text 8-fold. Modern models are likely much better.
I don't think many people disagree with this. The main problem is that labour has been what allows regular people to have negociating power with those who own most of the capital.
People are worried that if they lose this leverage, nothing is stopping the few who have most of the capital to just disregard the needs of the masses.
Not much info on the actual robot... For instance, I wonder how it has enough battery to follow a whale for 'months'? Which seems really unrealistic, as sperm whales can dive more than a kilometre, can't imagine an autonomous robot can support this kind of pressure, let alone for months at a time?
Why couldn't the calendar app expose in an API the read_calendar and update_calendar functionalities, and have a skill 'use_calendar' that describes how to use the above?
Then, the minimal skill descriptions are always in the model's context, and whenever you ask it to add something to the calendar, it will know to fetch that skill. It feels very similar to the MCP solution to me, but with potentially less bloat and no obligation to deal with MCP? I might be missing something, though.
Are you certain? My understanding was that this is automatically injected in the context, and in my experience that's how it worked. I never see 'ReadFile(claude.md)', and yet claude is aware of some conventions I put in there.
Afaik the experts are not usually very interpretable, and generally would be surprised if at least one does not change every token. I don't know what happens in practice, but I know at least during training, nothing is done to minimize the number of expert switches between tokens.
Personally defined <dtf> as 'don't touch files' in the general claude.md, with the explanation that when this is present in the query, it means to not edit anything, just answer questions.
Worked pretty well up until now, when I include <dtf> in the query, the model never ran around modifying things.
One important operation I've noticed in the examples that do end up with abiogenesis is having a 'copy' operation. In the bf version they use in the paper, one head can copy the byte under it at the location of the other head. Which makes it quite easy to make a self-replicator: just loop on the copy operation and move both heads, essentially (5 instructions). You could try adding the 'copy' operation to your setup and see if that helps!