If you turn things into individual discipline problems then surprise you get population level issues. As you pointed out people have other things to focus willpower on thats not this.

The fault, as it often does, lies in marketing. Turns out heavily marketed, hyper palatable food, designed to be minimally satiating so you maximally over eat is great for profits and terrible for obesity.

Anyway what's cute about this is that this isn't novel phenomenon. There are lots of parallels to this in other fields where often effective solutions do exist, but at a system level don't seem to work.

Telling people to diet doesn't fix population level obesity. Telling people about personal financial management doesn't stop people from accruing too much high interest debt. Telling teenagers to stop idolizing instagram influences doesn't fix body anxiety issues. Telling people to stop smoking/drinking doesn't fix addictions. 3-2-1 data backups absolutely work but people lose files all the time.

But modern bloat manages perfectly well to make apps that wait for network calls run poorly enough to give you a bad experience.

I'm not sure what a good mark would be, but considering this result lets put it at 2027-08-10 (One year from today).

I recently learned [0] that the USPTO makes it money from patents, its not government funded. Not only that but checking patents loses them net money while maintenance fees are the real cash cow.

The whole system is similar to the revenue model of a shitty journal that just publishes whatever research as long as the author pays. Except the office doesn't even need to care about their reputation in granting dubious patents because they have legal backing.

[0] It was a comment on hacker news, that I checked.

Non yolo mode where you manually approve each command is literally pure security theatre. No human on earth has the patience to discerne the huge stream of commands agents run.

So even if false negatives are technically zero, having 100% false positives is not acceptable.

So a fundamental premise is you have to have filtration. Where only very few things are surfaced for humans to look

The first option which a decent number of harnesses do is being able to setup an auto approve list. Like `ls` is fine, `cat` is fine etc.

Now maintaining this list is in and of itself a giantic pita. But the real issue is that it still has way too many false positives. Fundamentally it comes down to the halting problem where you can't really include important things like `bash python <So naturally the next best option is to use an LLM. Which isn't that stupid because even if its non deterministic at least it can dramatically reduce the false positives from the regex auto approve list.

All of this is of pretty minor importance though. You can't read as many tokens as a subcription can produce so more chat is not the value add nor super important.

I mean there are literally so many providers for free chat if you are willing to use several seperate apps.

The real value in these subs is using codex cli, much like the real point of anthropic subs is using claude code. Because agentic work actually does require a lot of tokens.

Maybe Luna efficiency gain was actually significant enough that putting all the free users and giving them super generous limits makes sense.

They might be doing this to improve the messaging of AI among causal users since right now there is a huge amount of datacenter backlash in the US due to AI grievances.

Maybe they have too much excess capacity or they really want to juice token numbers and market share on their dashboards for marketing.

I also wonder if being given access to an actually a decent model like luna with actual thinking budget instead of brainless "instant" modes will start to make causal users understand the real capabilities of these models.

It would be interesting to see something like:

$100 credits and $13 in taxes refunded as

$97 in refund after merchent fee + $12.61 as reversed taxes.

It's objectively better than nothing for sure. But the thing about this is such complicated nuanced solutions tend to get not done.

You can setup limits though to minimize loss. For example you can say auto reload $5 when your account hits $5 with an absolute monthly limit of $40. Then you don't have risk of losing more than the roughly ~$5 you have in an account.

But be extremely mindful when looking at anything touched by Anthropic or OpenAI.

These coding agents have terrible feedback cycles for lag and memory leaks. A particularly egregious one is GPU usage which is absurdly high in most of these apps due to terrible design and react render loop issues. Someone needs to make RL enviroments than punish this.

I will caveat this with saying being wasteful with resources is not new at all. But at this point it seems like no one is doing any amount of basic due diligence.

Can you figure out that 16 matrix vector multiplications into a concatenate is the same as concatenating first into a matrix matrix operation which can go on the GEMM. This is like the most basic thing you can imagine doing.

Turns out no it doesn't and there is a 23% performance difference by moving the concatenate up in the python code in my specific test.

Not to say that it didn't recognize it. I had an agent look at the XLA and the graph actually does a partial fusion into a sum and stack. But does not realize the whole thing is just a GEMM.

In more complicated examples the differences you can get can be much larger.

> this is not going to beat XLA for standard deep learning workloads. XLA has years of hyper-specific optimizations for linear algebra on GPUs and TPUs.

I'm not so convinced about this.

It's actually really easy to write two mathematically equivalent formulations in Python of something basic that have over a 2x performance difference in them after jax jit.

XLA is not that smart. And I'm not talking some niche nonsense I mean simple matrix multiplication graphs and residual connections on the CUDA backend.

Hours needed for prompt + Hours needed to check result + API costs.

You don't say "well let's add together the total yearly compensation of all the engineers and mathematicians at OpenAI that were involved" and throw that into the total cost. That's simply nonsense accounting.

The actual comparison you are making is some university researcher weighing between getting a grad student (several tens of thousands of dollars) vs typing up a prompt and sending a request to OpenAI for inference (as mentioned in the article, around $2000 in API and maybe a few hours for the prompt and harness).

Training the model is going to be amortized over other uses.

And yet this is the exact same company that has screwed up their android app so bad that the latex N^3 rendering problem makes it so having it explain it to me crashes the app.

Truly jagged beyond belief.

They should consider swapping this for a log plot.

I can imagine in 2027 academia looking like Moltbook.

Haiku was already in a ditch.

But this is coming straight for the jugular of a ton of models on openrouter.