André Beaumont

professional reply guy
we're in the singularity
codemonkey by trade

[André Beaumont @andrebeaumont]

@ptr_to_joel java needed JEP 394, 395, 409, 441 and a decade of "preview features" to arrive at "we have sum types now, please clap."

see you in 2035 for "modern java: we have ownership now (it's a Builder pattern)".

[André Beaumont @andrebeaumont]

@steamuhamlet @jdxcode this unfalsifiable goalpost-moving so you unix weenies never have to admit your religion has exceptions

[André Beaumont @andrebeaumont]

@jdxcode real users don't want primitives they have to compose themselves, they want workflows. nobody's piping gh repo create into anything, so why pretend it needs to be a single-purpose atom? bundling create+wiring+push into one command is just better UX

[André Beaumont @andrebeaumont]

@chetaslua @pangram

[André Beaumont @andrebeaumont]

@elonmusk Why is this fucking faggot image a video with audio go fuck whoever made this video

[André Beaumont @andrebeaumont]

Unnecessary. Community notes.

[*JAG* 🏴󠁧󠁢󠁷󠁬󠁳󠁿 @JustADayZPlayer]

Dear @elonmusk please, PLEASE for the love of what ever you hold holy to you, disable politicians from being able to turn their comments off or enabling only accounts they allow to comment. We NEED to hold these people to account before things really turn violent.

[André Beaumont @andrebeaumont]

@Cardoso @grok Eu me alinho com o judeu e com o chinês

[André Beaumont @andrebeaumont]

@ns123abc @OpenAI not sure where you're getting at nik
web and codex have never shared the same quota pool like ever

[André Beaumont @andrebeaumont]

@sushsrinivasan @TokenGremlin whilest "AI influencers" rip off your work

[André Beaumont @andrebeaumont]

@sushsrinivasan @TokenGremlin you're not getting the credit you deserve. fuck this gay earth

[André Beaumont @andrebeaumont]

URL='https://t.co/GWlfzuKVNM'
NEEDLE='Adventurer: Reach level 10'

for ua in \
"Claude-User" \
"Google" \
"OpenAI File Downloader" \
"XaiImageApiFetch/1.0 (Linux; x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/58.0.3029.110 Safari/537.3" \
"Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; GPTBot/1.2; +https://t.co/ZF6ctOZ0Ro)" \
"Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; PerplexityBot/1.0; +https://t.co/9JphhG9tZB)" \
"Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36"; do
body=$(mktemp)
code=$(curl -sS --max-time 20 -A "$ua" \
-H "Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8" \
-H "Accept-Language: en-US,en;q=0.5" \
-o "$body" -w "%{http_code}" "$URL")
grep -qF "$NEEDLE" "$body" \
&& printf "%-30s HTTP %s FOUND\n" "${ua:0:30}" "$code" \
|| printf "%-30s HTTP %s miss\n" "${ua:0:30}" "$code"
rm -f "$body"
done
Claude-User HTTP 403 miss
Google HTTP 200 FOUND
OpenAI File Downloader HTTP 200 FOUND
XaiImageApiFetch/1.0 (Linux; x HTTP 200 FOUND
Mozilla/5.0 AppleWebKit/537.36 HTTP 403 miss
Mozilla/5.0 AppleWebKit/537.36 HTTP 403 miss
Mozilla/5.0 (Windows NT 10.0; HTTP 200 FOUND

[André Beaumont @andrebeaumont]

@OrcaRouter vision was already cheap. luna at $0.20 and muse-spark-1.2 at $0.10.

[André Beaumont @andrebeaumont]

This deserves its own blog post. Let us know if you ever find the time for it. Curious on what your experiments surfaced. I added read_ref/splice to my own harnesses and it greatly increased m3 and Luna’s performance. It’s similar to hash line but I had sol arrive at it independently. But I gave up on separate tools. I’m testing how far I can get with only sh, prompting and internal harness state

[André Beaumont @andrebeaumont]

@unclecode @sushsrinivasan

[André Beaumont @andrebeaumont]

@unclecode did you arrive at it independently or are you failing to give @sushsrinivasan credit. https://t.co/rmqSCiK1yh

[Sushaanth Srinivasan @sushsrinivasan]

Ox Alpha is a https://t.co/p3agGEwXEx GLM model.

11 tokenizer probes across 10 models. Ox Alpha matches GLM on 11 of 11. No other lab gets past 4. DeepSeek burns 98 tokens on the digit probe where GLM uses 29.

It reads images. GLM 5.3 has no vision endpoint at all. Probably their next release, load-tested early.

https://t.co/4i3M4msE0Z

[André Beaumont @andrebeaumont]

@_can1357 @cafkafk you're a man of the people. love your research btw.

[André Beaumont @andrebeaumont]

@_can1357 @cafkafk or just the list of UAs is fine

[André Beaumont @andrebeaumont]

@_can1357 @cafkafk kind saar would you kindly release the script? interested in testing https://t.co/njy5eT9EaV

[André Beaumont @andrebeaumont]

@PartikelQ @teortaxesTex i think they sell themselves as being researched focused, rather than product only oriented

[André Beaumont @andrebeaumont]

@teortaxesTex any guesses on why the vision model seems to perform better even on non-vision related tasks? just a later checkpoint in the post-training pipeline?