This is crazy. What percentage of Anthropic employees are non-American? And what happens with internal models that aren't exactly Fable/Mythos? Does this actually affect any employee access to SotA internal models?
Who's Who of AI
Marco
Why they matter
Practitioner with public evidence across Compute & infrastructure, Culture, work & education.
- AI signals
- 173 past 30d
- Sources
- 4 distinct domains
- Discussions
- 11 past 30d
- Latest signal
- 11d ago
Language and keyboard stuff at Google.
I like computers and Korean and computers-and-Korean and high school CS education.
Georgia Tech → 연세대학교 → 東京工業大学.
Regrettably no longer based in Tokyo :/
https://theoreticallygoodwithcomputers.com/
What they're sharing
Statement on the US government directive to suspend access to Fable 5 and Mythos 5 anthropic.com
Last Translation Benchmark arxiv.org
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Retrofitting language models to operate over bytes - Nature nature.com
Articles & links
Marco reposted
Machine translation is not solved and it will take a while for it to be done arxiv.org/abs/2609.04173
AI Weekly's analysis
→
- Last Translation Benchmark ships peer-reviewed hard examples, each paired with handcrafted verification rules that flag concrete failure cases instead of assigning a metric score.
- LTBv1 covers submissions accepted before September 1st 2026, spans texts, images, audio and videos, and is designed to accept ongoing contributions.
- The authors argue MT is stuck: standard benchmarks are saturating, automatic metrics are gameable, and human evaluation lacks reproducibility and scalability.
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
- A two-stage 'byteification' retrofit used less than 1% of a typical pretraining budget, totaling 49.1 billion tokens across both stages.
- Bolmo 7B, retrofitted from OLMo 3, scored a +16.5% absolute improvement in STEM tasks over the earlier byte-level BLT 7B baseline.
- The method also converted Qwen3 8B into Bwen 8B and Llama 3 8B into Blama 8B, adding between 1.5% and 4.5% parameters.
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
- A new 'byteification' recipe retrofits existing subword language models into byte-level models with just 49.1 billion extra training tokens, under 1% of a typical pretraining budget.
- Applied to OLMo 3 7B, Qwen3 8B Base and Llama 3 8B, the retrofits (named Bolmo, Bwen, Blama) stay within ±5% of their source model parameter counts.
- Bolmo 7B records a +16.5% absolute improvement in STEM tasks over the prior byte-level baseline BLT 7B, with Bwen 8B improving on every category except GenQA.
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
- A new Nature paper retrofits subword-based LLMs into byte-level models using 49.1 billion tokens of extra training, which the authors call under 1% of typical pretraining.
- The byteified Bolmo 7B reports a +16.5% absolute STEM-task gain over BLT 7B, a byte-level transformer trained from random initialization.
- The two-stage conversion procedure is applied to Olmo 3 7B, OLMo 2 1B, Qwen3 8B Base and Llama 3 8B, producing Bolmo, Bwen and Blama variants.
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
- The paper converts existing subword LLMs into byte-level variants: Bolmo 7B, Bolmo 1B, Bwen 8B, Blama 8B, for under 1% of pretraining cost.
- Bolmo 7B shows a +16.5% absolute improvement in STEM tasks over the BLT 7B byte-level baseline.
- The two-stage procedure uses 9.8B tokens in stage 1 and 39.3B in stage 2, plus roughly 75M tokens of synthetic character data.
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
Marco reposted
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 buff.ly/l5ILoNT
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
Their own posts
Recent commentary
🚨 [Token][ization] Paper Alert 🚨 Tokenization is a wildly understudied area of language modeling despite it having effects across all of NLP. Over the past ~8 months, 32 (!) tokenizer researchers put together the most comprehensive survey of the field. Check it out!
"We've just been doing light abuse so far. We are trying to kick it up now." -OpenAI
The Claude Code use case that has saved me the most time so far is just "get this slightly out of date ML repo working on my workstation".
Requesting that everyone call me Dr. Parameter-Efficient Korean Character-Level Language Modeling for now on. I'm honestly so disappointed in spammers these days. No excuse in the age of LLMs.
I'm heading to @colmweb.org. Come find me to chat about tokenizers and Korean language modeling! I'll be presenting a poster at @tokshop.bsky.social about the need for a community tokenization benchmark a la the NanoGPT speedrun. It is meant to be a discussion, come share your ideas! #colm26
One funny thing about ChatGPT is that when I ask it some technical question and it starts doing the "searching the web" and then flashing through websites, it always does some weird ones like imbd. Idk what you are expecting to find there brother but you do you.
Why is there not a "I believe this paper is AI generated" button on paper review forms (that isn't shared with the author, but sent to the editors)? Ethical reviews are for ethical aspects of the paper, not ethical aspects of the writing process.
Their network
In Marco's orbit
Center = Marco. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you Marco? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/mcognetta-bsky-social)