woadwarrior01

TBF, Jan Marsalek (the Wirecard guy) is Austrian, although Wirecard was a German company.

FWIW, Ruuvi Air uses the Sensirion SEN66 sensor.

Main takeaway: foldl can be tail recursive.

> A bigger dense model beats it. Qwen3.8 27B ...

How is a 27B dense model bigger than a 78B MoE?

It's the other way around. Aging is the failsafe and cancer is the failure mode.

Isn't that just inpainting?

Not open weights yet.

The polyglot FFI feature (thanks to GraalVM) might be very useful for implementing "functional core, imperative shell" style designs.

Thanks for reminding me of it, I should really look into it.

Richard Gabriel's "Worse is better" vibes.

Same! I ported my free iOS utility app[1] to Android because a lot of people were asking for it. Just before I could publish it, they closed my Play Store account for a year of inactivity. Now, I haven't been able to re-create the Play Store account. Every time I try, money gets debited from my card, but the process ultimately fails. TBF, the money gets refunded after a month or so.

I'm happy supporting iOS, macOS and web.

[1]: https://cleanlinks.app/en

It depends. BTC is, ETH isn't.

This isn't really surprising. LLM reasoning and before that, chain of thought prompting are essentially forms of test-time compute scaling.

It takes one to know one.

There's a related Steve Jobs quote:

"When you’re a carpenter making a beautiful chest of drawers, you’re not going to use a piece of plywood on the back, even though it faces the wall and nobody will ever see it. You’ll know it’s there, so you’re going to use a beautiful piece of wood on the back. For you to sleep well at night, the aesthetic, the quality, has to be carried all the way through."

> For one, Apple was unwilling to have visible barcodes printed on the envelopes — but it wanted every step of the shipping and delivery of your card tracked. That wasn’t something the US Postal Service did. Apple and the printing company created an invisible barcode that was sprayed on the envelope, visible only under certain UV light, so that the envelope itself would remain unadulterated. And the USPS agreed to scan cards when sent, when processed at mail facilities, up to and including when they went out on a mail truck for delivery.

This is classic Steve Jobs. Sheer force of will.

aka Normalized compression distance (NCD). Its close cousin: Normalized Google distance (NGD) is also super interesting!

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

I have fond memories of using it for PIC16 as a teenager, ~25 years ago. It was a huge step up from writing assembly. My parents bought me a Microchip PICSTART Plus, but refused to pay for a Hitech-C compiler license. :)

That is such a canard, IMO. FWIW, Anthropic and OpenAI encrypt "thinking" token outputs in their models, while Chinese labs don't. If anything, it's more likely that everyone is using open-weight models in their synthetic training data generation pipelines. It's way easier to distill from logits than it is to distill from hard tokens.

https://x.com/EricSimons/status/2099252922098061714

Apple provides APIs for doing this in a privacy-sensitive way, instead of exfiltrating data like this, carte blanche using PostHog.

https://developer.apple.com/documentation/metrickit

This[1], I suspect.

https://docs.swift.org/swift-book/documentation/the-swift-pr...

Incidentally, there's a nascent OSS project called OpenSwiftUI.

https://github.com/OpenSwiftUIProject/OpenSwiftUI

If you care so much about user privacy, why do you exfiltrate events named "dictation_started", "dictation_completed", "dictation_rejected", "dictation_blocked", "insertion_failed", etc to PostHog, with properties named: "audio_seconds", "latency_ms", "duration_ms", "install_id", etc?

I spent ~10 minutes looking at it.

Vibe-coded website.

Both bundles contain quantized versions of the CohereLabs/cohere-transcribe-03-2026 model. The model in JEXXA.dmg is MLX int8 g64 affine quantized. JEXXA-Small.dmg has the same model but MLX int4 g32 affine quantized. There's also an int8 quantized WeSpeaker ECAPA-TDNN speaker-verification ONNX encoder. No acknowledgments for both models (doesn't the Apache 2.0 license require it? WeSpeaker's cc-by-4.0 license certainly does). Both bundles also contain full-blown Python 3.12.8 runtimes with about a dozen packages installed.

Apps are unsandboxed menubar apps and also contain PostHog analytics and Supabase auth. So, I wouldn't run it on any of my machines.

A few engineering hygiene issues like a .pytest_cache directory, Python code, .DS_Store files, etc.

Given the above, I suspect the "I am training better models." is just marketing speak.

nb: The demo link without LinkedIn tracking slop: https://www.linkedin.com/posts/sankyde_jexxa-demo-httpsjexxa...

Very interesting!

I suppose the next logical thing to do would be to train a language model to generate random flags, call it Artificial Flag Intelligence (AFI) and raise a $10m pre-seed at $100m post. :P

I'm not defending claudeslop, but TBF, anaphora and its lesser-known sibling: epistrophe far predate the English language, let alone something new like LLMs.

On a related note, I came across this USB-C KVM on pre-order the other day. It ostensibly supports Tailscale, which would be really useful.

https://sipeed.com/nanokvm-go

This one's authored by a human, the other one is authored by Claude.

> Does the M4 and later ANE expose any additional capabilities, or is it just a higher-performance iteration of the same thing?

IIUC, M4 introduced a fast path for INT8 weights and activations (w8a8). M5 Ultra, M6 and A20 have two ANEs.

> As an aside, the introduction to this article seems to conflate the ANE with the Neural Accelerators (NAX)

Yeah, that part is true. NAX cores are matmult accelerators, closer to tensor cores in NVIDIA GPUs.

TBF, what difference does it make when the very harnesses you're using these "hacks" in are themselves vibe-coded?

They have a history of making grandiose claims like this[1] from 2024, with no visible products or research.

[1]: https://magic.dev/blog/100m-token-context-windows (also linked to in their blogpost)