You should add features to address how we actually learn: Here are the 4 "Horsemen" of sticky learning.
1. The Generation Effect
The Science: Your brain learns better when it attempts to solve a problem before it is shown the solution—even if you guess wrong. The struggle to "generate" the answer creates a cognitive hook that the correct answer can latch onto later.
2. Retrieval Practice
The Science: Re-reading puts info in. Retrieval forces you to pull info out. Every time you retrieve a memory, you modify it and make the neural pathway stronger. The MIT study showed AI users failed because they never had to retrieve anything—it was all external.
3. Spaced Repetition
The Science: Cramming works for 24 hours. Spacing works for life. You need to let yourself slightly "forget" information before retrieving it again. This effort to recall faded knowledge signals to your brain that this info is vital for survival.
4. Interleaving
The Science: Traditional school teaches "AAA BBB CCC" (Block practice). Real life is "ABC BCA CAB." Interleaving mixes up different types of problems/subjects. It forces your brain to not just execute a solution, but to first identify which solution is required.
You can decide how best to incorporate these elements in what you built, but without guided practices that incorporate these practices, it won't solve the "we forget what we listen to" problem of learning through podcasts
Hey bro, been working with the Godot MCP for about a year now myself. This is really cool! The two loops I think you should add / address: 3D assets / rigging. This is the hardest thing to do in the current Claude Code to Godot MCP loop right now. Claude can reliably make 3D assets using the in game creator, but ideally you have a system that works with external models (or better yet, generates them ala something like spline or similar system in Blender) and then can rig them up in game. 2. Expanding the evaluation loop from just screenshots to actually playing the game. I just started building a game in Claude Code Desktop, and the Claude agent there can actually PLAY the game. That's a huge unlock. I've been thinking about how to get this same functionality in the Godot MCP server but if your system can do it that would be awesome too.
This is the defining divide of AI, period. Whether you're a craft lover of art, writing, music, code, hell, business processes and the idea of "doing work." There are those who love the craft, and those who want the result of the craft. AI is a faster path to that end result (whether you're happy with that result is another matter). From that POV, it could lead to us speed-running our civilization into another era; abundant prosperity, or full on collapse. Bro...
Okay, I'm tired of reading the debate about costs going down and therefore Ed is wrong. The cost of running the inference is not the problem. The cost of the input CHIPS is the problem. Let's return to Dario Amodei's interview [0] with Dwarkesh, shall we, for AI Economics 101?
Here goes:
The Epoch data everyone keeps citing measures the price per token charged to API customers. That's the sticker price. It tells you nothing about whether the business is viable, because the existential risk for AI companies isn't the marginal cost of running a query. It's the upfront capital expenditure on chips and datacenters, committed years before you know what demand looks like.
Anthropic CEO Dario Amodei spelled this out in his Dwarkesh interview. Here's the short version:
1. Data centers take 1-2 years to build out.
2. Each gigawatt costs roughly $10-15B per year.
3. The industry is currently at ~10-15 GW, scaling roughly 3x annually.
4. By 2028, ~100 GW. By 2029, ~300 GW.
5. We're talking multiple trillions per year in committed infrastructure spend across the industry.
Now NVIDIA's Q4 earnings [1], which printed today:
1. $68.1B in quarterly revenue, $62.3B from data center alone.
2. Full-year: $215.9B, up 65% YoY. Guiding $78B next quarter.
3. Someone is writing those checks. Those checks are not refundable.
Dario, who believes we're 1-3 years from a "country of geniuses in a data center," described his own demand prediction as a "hellish" problem.
His exact framing: If this revenue comes in at $800B instead of $1T, "there's no force on earth, there's no hedge on earth" that could stop him from going bankrupt if he'd bought compute at the higher projection.
He's at ~$10B annualized revenue today, and he won't commit to buying at the scale his own thesis demands, because being off by a single year is fatal.
This is the actual argument (I'm not saying this is Ed's argument, but this is the argument against these companies). Not "inference tokens are expensive."
The argument is structural: these companies must pre-commit billions in non-recoverable CAPEX based on demand projections that are, by the CEO's own admission, a coin flip.
The gross margins on serving tokens might be great. But the training spend for next-gen models grows exponentially, and it has to be funded before that model earns a dollar.
The Epoch chart measures what customers pay per token. It doesn't measure the $215.9B NVIDIA invoice those customers collectively funded this year, or that these chip purchases are one-way bets against future demand that may or may not materialize.
Inference costs going down 20x is wonderful for consumers. It tells you almost nothing about whether the companies making those chips, or the companies buying them, will survive the demand prediction gauntlet.
And if we're being honest, the Epoch data showing 9x to 900x price drops per year should make you more nervous, not less, because it means the asset you bought last year is depreciating at a rate that makes used cars look like gold bars.
This is where I've landed as well. One caveat: hard to say if anyone's revenue numbers outside of those two names is reliable. Perhaps anyone who isn't the big four is at risk of bubble trouble.
