jw1224

An LLM takes (text in) -> (text out).

Jev takes (…questions in) -> (…probabilities out)

So Jev won’t write a story or emit arbitrary structured data. But if you ask it the right questions, it can make near-instant “decisions” against those questions, with accuracy and world knowledge on par with LLMs. The economic advantage is that it’s parallelizable and can give back up to 255 answers at once, in milliseconds.

With a ratio of 1:√2, you can cut the sheet in half parallel to its shorter side, and each half keeps the same proportions.

There are some amusing apps which read the MacBook's lid angle sensor. I remember one of them played a creaking noise through the speakers when you open or close it... Somehow I doubt App Store review will let that in for the Duo, but we can always hope.

I've used a similar technique with my own tools. It helps to save context overall, by deferring the detailed instructions not only to when they're needed, but also to where they're needed.

Simply loading the browser Skill alone, doesn't always mean a browser is actually necessary (or supported). The user might've want them to control their main desktop browser; or the user might be pointing them toward a read-only HTML file that doesn't need rendering; or the sandbox might prevent network access; etc...

Codex is especially 'keen' to use Skills (especially by comparison to Claude), I often see it load Skills which it doesn't go on to use, because the main Skill content says "when not to use this". So I'm sure this is a product design decision to help with those kinds of circumstances.

Interesting, I've come to the opposite conclusion: a lack of types (or types that are only weakly enforced) costs me significantly more tokens in the long-run to maintain, and makes it far too easy for models to silently introduce bugs.

I run all my projects now in TypeScript with the strictest possible settings, including disabling `ts-ignore` markers.

(This would drive me absolutely insane, but my agents get over it pretty quickly!)

> When Claude Fable 5 declines a request, the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response, not an error

This is precisely what comes to mind when I think “successful”.

Not OP, but I generally agree. Models are powerful enough now to reliably instruct other models. They don’t need fancy tools or IDEs, just the command line.

With deterministic workflows, type-safe languages and test suites, agentic loops pretty much “can’t fail”. They will continue until the types resolve, the tests pass, and the project requirements are deterministically met.

By that point it’s literally just a case of typing a prompt in to a text field, and waiting.

Skills already are "just asking the model". Unless you'd prefer to type out the same instructions every single time?

Skills are literally just Markdown documents that get loaded into context when the /skill-name is invoked.

MLX is Apple’s own machine learning framework, designed for Apple Silicon: https://opensource.apple.com/projects/mlx/

> “To stave off some obvious comments:

> yoUr'E PRoMPTiNg IT WRoNg!

> Am I though?”

Yes. You’re complaining that Gemini “shits the bed”, despite using 2.5 Flash (not Pro), without search or reasoning.

It’s a fact that some models are smarter than others. This is a task that requires reasoning so the article is hard to take seriously when the author doesn’t even turn reasoning on (nor mention it anywhere else).

I asked the exact prompt to GPT5 Thinking and got an excellent answer with cited sources, all of which appears to be accurate.

The “full Tailwind experience” is already freely available. What “lost opportunities for deep integration” is a frontend CSS framework missing?

Tailwind Plus (the commercial product) is like buying an off-the-shelf template. It’s just a collection of themes and pre-built components — useful for devs who want to get started quickly on a project, but it’s cookie-cutter and can easily be replicated by anyone with Tailwind itself.

Associative arrays are iterable internally without needing an interface or key tracking. You can just foreach them.

> Let me guess.

This just proves my point, you call this a “hot mess” whilst completely misunderstanding what it’s even used for.

The Iterator interface isn’t even used in the comment you first replied to.

I don’t think you understand the purpose of PHP’s Iterator interface.

I’m not sure I follow, what exactly is your complaint? The Iterator interface is described as:

> Interface for external iterators or objects that can be iterated themselves internally

Note “external iterators or objects”. The Iterator interface is not exactly everyday PHP, it’s a specialist utility for making classes iterable so they can be accessed like arrays. Most developers will rarely use it directly, and it’s not being used in the parent comment’s example either.

Iterating over something requires knowing where you are in the sequence, so of course you would need to implement a method to get the current position of the iteration.

PHP doesn’t force keys... You can omit the key and simply write `foreach($items as $value)`

You can buy phenol-free thermal paper, it’s about 20% more expensive where I live but much safer for you, and the quality is just as good.

Great first article, and very interesting to see someone else using a receipt printer for bite-sized task management!

I have a variety of automations running which print actionable tasks to my receipt printer via a Raspberry Pi. It’s nice having a real-life ticket I can take hold of.

One thing to be aware of if you’re handling receipts frequently: make sure to buy phenol-free thermal paper. Phenol is toxic and some types of it are banned in certain countries.

Well if that hypothetical situation ever happens, you can just switch to Linux then.

If that’s the one reason, have you considered just… not using the AI features?

I thoroughly enjoyed reading this. I peaked at 137wpm[0] and I’m not ashamed to say I love typing long sentences. I type fast enough that my thoughts normally can’t keep up with my fingers, but when I plan a sentence out in my head, the satisfaction I get from watching my inner monologue transmogrify into words onscreen is palpable. It’s a rush. And I totally get it. Great article!

[0] https://data.typeracer.com/pit/profile?user=mavis_b

> If we […] tried to use these systems to solve the kinds of problems we need Einsteins, Hawkings and Taos for, then we would be in for one miserable disappointment after another

We can literally watch Terence Tao himself vibe coding formal proofs using Claude and o4. He doesn’t seem too disappointed.

https://youtu.be/zZr54G7ec7A?si=GpRZK5W1LDvWyBBw

This is some excellent trivia. Thanks!

Do you not think the AI output looks far more polished and print-ready? Canny edges have a lot of noise and don't look at all clean for coloring book purposes.

Are you sure about that? It correctly solves the puzzle.

Each step (though a little confusing to parse) does make sense in context.

Even Sabine Hossenfelder did a video about it today too: https://youtu.be/dOIlMdIqXbQ?si=FG4ZdDq6jkjk_nMp

Abstract:

> The maximum element of the vector output by the Softmax function approaches zero as the input vector size increases. Transformer-based language models rely on Softmax to compute attention scores, causing the attention distribution to flatten as the context size grows. This reduces the model's ability to prioritize key information effectively and potentially limits its length generalization.

> To address this problem, we propose Scalable-Softmax (SSMax), which replaces Softmax in scenarios where the input vector size varies. SSMax can be seamlessly integrated into existing Transformer-based architectures. Experimental results in language modeling show that models using SSMax not only achieve faster loss reduction during pretraining but also significantly improve performance in long contexts and key information retrieval.

> Furthermore, an analysis of attention scores reveals that SSMax enables the model to focus attention on key information even in long contexts. Additionally, although models that use SSMax from the beginning of pretraining achieve better length generalization, those that have already started pretraining can still gain some of this ability by replacing Softmax in the attention layers with SSMax, either during or after pretraining

Overall I’m extremely impressed.

I never thought I’d say that about any kind of government policy, but here we are...

It’s decisive, forward-thinking, and sees the bigger picture. Obviously, the proof is in the pudding, but this is a very promising start.

Is this something you’d consider sharing? I know many of us would find it really useful!