amoss

An interpretation that is obvious can be easy to arrive at - but that does not make it correct, only low effort. If the test had not failed then it was not ambitious enough. If they had not pushed the rocket to its current breaking point then they would not have learned as much about it, or its current modes of failure.

Sometimes it is worth looking past the obvious interpretation. I thought the article was a good read, kudos to the author.

If it is based on an LLM trained on the internet then it is probably outputing "gib moar srvrs plz" every twenty seconds.

Thanks for the explanation - interesting link. The same sort of rigidity appears in functional programming styles, where is bound more tightly into the data definition. I'm thinking about the sum types in Haskell for example, where the programmer can make a rigid definition of the boundary that you are describing and then it is reflected in the structure of the code that manipulates the types.

There is an algebraic similarity between Horn clauses, BNF grammar clauses (if you consider concatenation and choice to be operators) and sum types. After working in logic programming, functional programming and imperative styles I would agree that the rigid definition part makes it easier to write the code that will sit on top, even if that code is far less rigid.

There is some kind of bottleneck shown in Table 4 and I can't see what it is by reading the text in section 7. The IPC/core drops from about 2.5 to 1.5 for the same stall rate at the number of cores is increased?

What connection are you drawing between the two? BNF is used to define language grammars. Prolog is written using Horn clauses. There is not an obvious connection.

Does your model explain the need for a minimum wage?

* People like to eat.

* Capitalists don't like to give people money.

* People will not work for free.

Just about covers it.

For the ongoing maintenance costs I have assumed that the costs per GW are given with respect to faceplate capacity. You have assumed (or know) that they relate to the useful output. I went back to check in the source document (and in reference 41 where they originate from) but could not see a definite statement either way. Did you find something that clarified this?

How many liquidity injections and financial bailouts we will need over the twenty year timescale in the plan? One advantage of this plan is that it provides a direct liquidity injection and provides demand in several key industries. It goes without saying that an EV company would be a huge beneficiary from some of those funds, but I wonder what the net cost would be if it offsets the need for even some of the trillions being used to glue together the world economy since the events of '08.

There are two factors that you have not addressed:

- Ongoing maintenance costs (the numbers are in table 5, but essentially wind requires 2x the maintence cost of solar on shore, and the ratio is worse off shore).

- Material used in construction (table 15), the amount of concrete used is 6x greater for wind than solar. This implies a far higher CO2 cost during manufacturing. The steel used is far less for wind, I didn't see a complete breakdown of CO2 emissions for manufacturing, say per/GW of capacity.

Your overaching point seems to be that solar is not an entirely better approach than wind. But the suggestion in the plan is a mixed approach using both solar and wind. One point for comparison (made in a sibling comment) is that solar allows a more distributed approach, and a more incremental roll-out. So let me flip your point around somewhat and ask the converse: why concentrate more on wind instead of trying to optimize a balanced approach over all of the variables?

Perhaps this is more what you are looking for: throw together a playlist with a particular "vibe". Doesn't have to be large, 5-10 tracks works fine. Either switch on enhanced and look at the suggestions, or use the separate section for suggested additions at the bottom. Add the ones that fit the vibe, reject the rest. Repeat. I've found that it mines niches particulary well and discovers new tracks much more effeectively than the pre-canned discover playlists, or daily suggestions.

uptime 8:26 up 38 days, 9:49, 4 users, load averages: 1.59 1.51 1.60

The number of tokens in the training set is not a good measure of the information content of the training set - if it was then the training would not work as the training set would be statistically random.

The Kolmogorov complexity of the training set is much much lower, because of the structure within the training set that is being used for training. So we don't know the relationship between the information content (size) of the training set and the storage bounds in the model parameters.

> It amazes me that it understands spatial layouts from training only on text such that it can draw physically accurate diagrams, and it understands graphics libraries well enough to draw with them.

Is there evidence of this? The Whitehouse floor plan is very well known, and available online in many different formats and representations. Transforming one of those into a sequence of calls would be easier.

Have you tried this with a textual description of a building that does not have any floorplans available, i.e. something unique?

That's interesting. What kind of problems?

- autonomously design, code and distribute whole apps (but not the most complex ones)

This is a bold claim. Today LLMs have not been demonstrated to be capable of synthesizing novel code. There was a post just a few days ago on the performance gap between problems that had polluted the training data and novel problems that had not.

So if we project forward from the current state of the art: it would be more accurate to say autonomously (re-)design, (re-)code and distribute whole apps. There are two important variables here:

* The size of the context needed to enable that task.

* The ability to synthesize solutions to unseen problems.

While it is possible that "most complex" is carrying a lot of load in that quote, it is worth being clear about it means.

Not to disagree, but you seem to have skipped 0) Increasing the parameter size of LLMs is a path to sentience.

Ah, ok so you are considering the latency rather than the throughput. I was thinking that being able to perform the update within 16.6ms should not be too hard - there is a sibling comment about SSD latency but the working set should fit in RAM.

Sure, so there will be latency, maybe 4-5 frames with input lag. But this is normally bearable in a game so I'm guessing that it would work fine for an office app. It would be difficult to remove, but 80ms response time from an application that can always hit its 16.6ms deadline should be workable.

Why?

In 16.6ms a modern processor will execute about 128 million instructions. We need to locate about a screenful of text and/or cat pictures. We need to perform layout and then render it. The amount of work being done is not very much, or very difficult. The resources available are luxurious.

What part of the problem is it that makes it an unrealistic target?

https://archive.is/N7xvO for what it is worth.

But, tl;dr Tories are gonna Tory.

This assumes that the LAN is the bottleneck. Gigabit ethernet tops out at 120MB/s, which is about the speed of spinning rust on a NAS.

It can fix the planet or it gets the hose again.

I wonder how this will apply to code that comes out of systems like copilot?

lol, of course. I hadn't realised it is so long since I wrote any Haskell...

The interesting part of the exchange is the contrast between the surface level syntax of the exchange and the semantics below. It reminds me of teaching in a lot of ways. The level of polish on the immediately apparent surface level of the conversation is very high: reasonable tenents, good justification and ability to explain them.

But then applying them directly into an example falls down a bit. The language is a bit of a mismash and parts of it might not work semantically. The use of actors looks a bit out of place, and raises a lot of questions about how the memory management and scheduling is working under the hood. Despite the bold aims in the tenents thre is nothing in the language exposed so far that addresses them: why would the language produce more modular code in the large? what features of the syntax would increase the level of abstraction in code written? etc etc

Overall this would warrant a (barely) passing grade by modern standards, i.e. it can regurgitate the material succesfully but without demonstrating any deep understanding. I find the result quite interesting as it demonstrates how much of what we judge to be quality is actually fitting a particular pattern of response (w.r.t a specific field of study) rather than exploring the material more deeply.

I like the use of the arrows in the function definitions though :) I've seen that somewhere before and will be "appropriating" it for the project that I'm currently working on.

This is true, but not currently relevant. The cost (difficulty) of training does not indicate how good a fit the model is for copyright protection.

I suspect that part of the arguing this out in court will be making those economic arguments as part of an attempt to extend copyright protection to model weights - but that is a separate issue from if they are protected currently.

Part of the economic argument for copyright is to encourage the investment of time into producing new works, so it may be a persuasive argument for extending protection. Although (relatively) recent history suggests that persuasive argument is not as powerful as giant wads of cold hard cash in extending copyright coverage.

That is the point that I made - a baseline for comparison is not defining the method that produces the best understanding of the code. It is defining the worst that we would expect somebody to try.

A baseline to compare against would be googling the name of the function and reading a random result page instead of the actual code. The OP is demonstrating that the system does not rise above that baseline.

Those protections rely on "systematic arrangement", i.e. that some effort has been put into finding, verifying and collection a set of facts / individual pieces of data. It is not obvious that would apply to the set of weights that come out of a training algorithm. (standard I'm not a lawyer, I don't know but I suspect it is unclear enough that it would need to be argued in court).

It helps if you take a deep breath.