Who called them hallucinations and not ÆIslop's Fables

“The point about formal notation is *not* to output it… it is a tool for thought. The point about LLM generated [code] is that you ideally don’t [formally notate]. Saying the LLM could output [formal notation] is missing the point” ✍️ Transcript (also in alt text): gist.github.com/lmmx/9838843...

Big day for ominous greyscale photomontageurs

Today I learnt about "bypass charging", a phone feature where it can run directly off the mains without using the battery at all. No battery charging, less heat, less battery wear. Nifty. www.androidauthority.com/phones-with-...

Opus 4.6 flubbing a dedup task baddd

They're turning the AIs into golems now

hooooollllly.... this is so cool, i did not expect it to actually work.

Currently: vibe Viterbi’ing

Electra: underperformed vanilla BERT (and BERT cased seems to learn the task faster, though still reaches the same limit of F1 which I can only interpret as my data limit showing) I feel like I’m being pranked wdym BERT is SotA

Fun fact I learnt today: a method for data augmentation in NLP is ‘back-translation’ (translating to another language and back)

L + Ratio + Cope + Seethe + Send + Sync + 'static

am I allowed to question minibatching received wisdom or

becoming a retvrn guy for WordPiece

Has anyone done like… full source code recording in training instrumentation metadata? not just commit hash 💭

Yelling at the loss like it’s TV

Saw ModernBERT loss shoot down straight but plateau much sooner, F1 not reaching as high as plain BERT Caused by tokeniser misaligning BIO labels w/ subword tokens, so model converges on predicting mostly O tags and achieving artificially low loss while failing to learn the actual seq labeling task

deBERTA v3 blew up on me in the first epoch it's transfover

Training models all day and night as God intended

*Shakes the pipeline through several more rounds of simplication*

making the most beautiful pipeline… ty takes you so much further than flake8 alone

All this time I never clocked the Pydantic ConfigDict is what the dict has to conform to, not necessarily instantiate! 😅

TL seems to have thoroughly turned on CC

I guess thematically last year I did a lot of work to be able to verbalise data structure and then to operationalise that through genson-cli, and this year I’m working to verbalise flow To say anything seems like you need to decide your rules to abide by and prove why they’re operationally valuable

Concept: LLM as prion disease for software

🏞️ cog.spin.systems/model-distro...

Unsure if my realisation about making data sources immutable in model training pipeline code is completely trivial or totally off-piste (“hardcoding corpus version specifiers in an immutable data registry = good, input source should not be runtime configurable”)

*checks pyreadiness.org/3.14/ *

Selecting between paths based on their existence on the filesystem using default validation at runtime gist.github.com/lmmx/0a7c215...

and lo, wend is born 👣 pypi.org/project/wend/

You are experiencing code casinofication by vibe code lottery I’m experiencing it by guessing which four letter names will be sufficiently distant to existing PyPI packages’ we are not the same 🎰 🎰 🎰

Turns out the electricity transmission network looks pretty cool when animated in Blender ⚡ The idea is the animation shows flow direction along each line, though right now the directions are completely made up until I get my head around PyPSA.

Nice, Modal got uv sync support in summer (taking groups, extras & frozen) modal.com/docs/referen... uv workspaces not supported/roadmapped PR: github.com/modal-labs/m...

grepow! 💥 pypi.org/project/grepow/ 📂 github.com/lmmx/grepow

kicking myself for somehow still not internalising you can do mv with braced path sets llike /path/to/nested/{src,dest}.py

Concept: grep.app local CLI which sparse-checkouts each repo with a cone over just the hits

gas soundcloud.com/teknaa/dolly...

User has been bitten by the model training mania

No better feeling than all the data artifacts falling into their rightful place

User has acquired push access to prek 😄🕺🪩

Surprisingly Claude *cannot* teach you Helix (vive les forums!)

The urge to forge the Platonically perfect pipeline Something like: the t-string pathlib idea I blogged about recently+ versioned artifact layout + I/O checking guards... etc etc

TODO: check if nothingburger github.com/github/spec-...

oh that? (getting visibly unsettled) that's my pet cattle

LLM """health files""" declared stale on arrival on day ending in y

Feeling like if I can conquer my various inhibitions around using Modal specifically I can become deeply actualised (somewhat paralysed by the power of the machinery) An LLM is a lot like an e-bike, the engine is v powerful (hence passivating), if used carelessly you’re gonna be slammed into a wall

Sooner or later someone has to come up with a cross-cutting response to generative AI shredding social trust, rather than all these ad-hoc one-offs that special case each domain viz. OSS social coding remedies (disclosure, deposits), conference reviews (reviewer/ee mutuality) …

...horrifying message without context, thank you

Running Modal for the first time… very… carefully

Just turned a few 000 doc corpus into a nice parquet artifact using some generated TOML manifests (which Claude had tried to overice) rather than mapping via code globals Straightforward task but I felt more in control because the preconditions were explicit semi-literate artifacts rather than code

sqlite is overrated… it’s alright we can say it