_aavaa_

It could be a study of 2.5B children and it wouldn’t make a different for the people peddling this nonsense.

The whole thing started with fabricated research, it’s lies all the way down.

> the visual difference is directly exposed to the player. A sixth finger on a hand is an issue in the way that as line of dead code isn't.

But a line of buggy code will be substantially more impactful than a poorly drawn model. An extra finger on a model doesn’t cause your save to be corrupted or a quest to hard lock.

I’d go a step further. The game code is the game. Everything else is there in service of the interactive experience.

Then based on your own analysis, at least the electron bloat angle (that many people, at least many developers, hate and wish was native) has nothing to do with the low level knowledge of the developer. And I would argue that a good deal of other areas of slow software will have the same conclusion, i.e. the slowness is not due to lack of knowledge.

I don’t think that lack of knowledge about memory training is a hill to die on when it comes to causes of app bloat.

The people who need to most think about performance are those making the frameworks and languages. Why are bloated electron apps so easy to make but building native cross-platform apps with similar ergonomics hard? Why are people choosing electron?

But we’re not talking about recalling how to solve a particular problem, or a fact about an algorithm, we’re talking about API names and syntax.

I don’t have to remember assembly intrinsic since the compiler deals with that, I just describe the higher level goal and it implements it.

Yes, now compare it to modern EVs or non-EVs from every other brand.

I too hate this practice, but let’s not false for the anti-China rhetoric that makes it sound like what they’re doing is even remotely out of the ordinary.

“Computer add a filter that removes all duplicate entries from this data structure. Two entries are duplicates if they every value in their rows are identical.”

This is the same thing as pressing the power button on my computer; I want the thing turned on, I don’t want to have to know about memory training.

> before AI, it was a valuable skill to be able to recall APIs quickly

That’a the low value they’re talking about. And all of that stuff is accidental complexity.

To use their example, you need to know that dropping duplicates is a good idea (and why). Knowing what the syntax and api calls to do that is, in my opinion, unnecessary.

How much would I get paid?

I disagree. The framing of focusing on primary energy and then sprinkling efficiency afterwards has two issues: 1) it pulls focus away from the thing we want (the end result) and 2) it makes it much easier to misunderstand (or to misrepresent) because the efficiency differences can be omitted.

And graphs like the one comparing wind turbine output to petrol car consumption are inherently deceiving. Two values are put side by side with the same units and then talk about directly as if they are comparable. But they simply are not. A kWh of chemical energy and a kWh of electricity have as much in common as a US dollar and a Jamaican dollar.

I wonder if people would use that phrase as much if they understood where it comes from.

https://en.wikipedia.org/wiki/Shouting_fire_in_a_crowded_the...

I’m not saying he was unaware of heat pumps, I’m saying he’s comparing apples to oranges simply because they use the same units.

Look at “3 - Cars” on page 29. He says the typical car uses 40 kWh/day. 40 kWh of what? Chemical energy in the gasoline.

The go to page 33 where he looks at how much energy onshore wind could produce per days in the UK. His number is 20 kWh/d. 20 kWh of what? electricity

He then compares those two numbers directly and uses that comparison as the basis of his arguments: “Britain’s onshore wind energy resource may be “huge,” but it’s evi- dently not as huge as our huge consumption.”

This is simply incorrect. A combustion engine converts less than half of the chemical energy in the gasoline into mechanical work that can move the car. The electric model converts >90% of it. So we don’t have to replace 40 kWh/day, we have to replace less than half of that since the electric process is more efficient.

This same issues, the primary energy fallacy, underpins large parts of the book.

The books is good(ish) for it's time, but some of its analysis and forecasts are fundamentally flawed since it falls for the primary energy fallacy by comparing the chemical potential energy (in J) directly to electrical energy (also in J). The two are fundamentally different things and called be compared 1:1. E.g. To heat up your home with natural gas you need ~1J of chemical energy to get 1J of heat into the home, but with a electric heat pump you only need 1/6 J to get 1J of heat.

It is also a product of its time in terms of wind/solar vs nuclear. His forecasts of the impact of solar and wind is based on prices and performance from 2008. Prices have come down an order of magnitude since then, and performance and lifespan have increased drastically.

Like front-running Millennium prize solutions?

If you don’t trust and are on the lookout for silent nerfing, how are you protecting against it detecting a benchmark and using the un-nerfed model?

Yes, I fully believe that's why they're pushing on this. But as the general population, we don't need new laws to protect us from this AI-related problem. We just need enforcement of what is already on the books.

Why do we need new regulations? Hacking into hugging face is already illegal.

Is there really no limit on number of concurrent agents on either OpenAIs or Anthropic's plans?

That works for you personally, but what are people going to do 20 years from now? Fight over the dwindling stock of 40 years of vehicles?

To say nothing of the tradeoffs that come with an older vehicle (safety and pollution).

“Ahh, and yet to participate in society.”

This is a direction all manufacturers are going into, so there is little actual choice. And buying used is not a real solution; where do tomorrow’s used vehicles come from?

If the model provides me with bad results because it's dumb, I don't care how quickly it does it.

Are we prepared to play kingmaker and pick one SMR design to standardize on so that those can ride their theoretical learning curve (assuming we get enough buyers)? Because politically that doesn't seem likely, and I think we'll end up with the units that do get contracted to be spread across multiple firms, none of which make it out of the expensive first few units.

> Even Flameville 3 in France was significantly cheaper than Hinkley point C

Per wikipedia: initial cost estimate €3.3 billion, cost in 2020 (6 years before it was done) €19.1 billion, so 5.8x the budget. Supposed to enter operation in 2012, entered in 2026 "only" 14 years late.

These are terrible results.

That EPR2 series are "expected" to cost less says absolutely nothing. Flameville 3 was expected to cost 83% less too, and I except that if they swindle people for the money to build it we'll see it'll be just as much of an overrun.

The nuclear industry seem absolutely incapable of sticking to a timeline or budget. And any one of these failures should have been disqualifying. But they get to do it over and over again.

It will reliably cost only 10x, and it theory it will roll of a production line.

The AP1000 reactors at Vogtle were explicitly marked for the modular construction as a key selling feature, which would reduce time and cost. It came out 20B over budget and 7 years late.

After the two previous America themed rebrands, I'm really surprised it's not American Intelligence.

That’s not saying much. The selling point of pixels has always been the software.

> Article says that there's no human intention or design directing the output we get

I mean that’s objectively wrong for any model using RLHF.

If you are using their coding plan for coding, then yes you can easily hit such cache rates, with a good harness.

I’m getting 97%.

I'm not defending their actions, but we should be clear about where the law currently stands: Anthropic was found to infringe because of the torrenting, not because of the training.

Their plans are still worth it if you use their models. You can see how many tokens you can except to get based on plan here: https://docs.z.ai/devpack/overview#estimated-token-allowance

The max plan will provide ~1,100 USD of GLM-5.3 or ~260 USD of GLM-5.3-flash per month for 168 USD. I can personally attest to these numbers through omp (~97% cache hit rate).

Unless you are able to highly parallelize (your work, you won't be able to hit your hourly or weekly quota using the flash model simply because it's so slow.

They give you ~3x more flash tokens, which maybe comes out to ~2x more actual work after accounting for the extra thinking it does to achieve the same result. The mental model, for not getting angry, is 5.3 is fast mode by default, and you can disable fast mode for 2x the work output at 1/3-1/10th the speed.

They're serving me 5.3 at ~40 tok/s and 5.3-flash at 30 tok/s (according to omp).