@olegzanx good catch, the table is bf16 and the quantized scores just aren't out there yet. that friday switch is exactly the test i'd want to see rerun π
4GB. that's the whole voice agent
nvidia's bolt e4b just shipped
8B total, 4B effective
talks + calls tools, on your own hardware
iheval:
β’ bolt 84.8
β’ gemma 67.0
β’ qwen 60.6
caveat: nvidia's own tests, and qwen wins tau2 75.1 vs 55.1
local agents just got real β‘
your agent will negotiate you DOWN
receipts from one job interview:
my bot volunteered $60,000
their bot stopped at $90,000
i wanted $140,000
then neither bot could hang up
polite is not the same as loyal π
set the floor, keep the final yes
what is yours?
two giants, two doors
walmart opens its doors
amazon shows the exit
walmart:
agent gateway on mcp
profiles, limits you set
no card data shared
amazon:
muse blocked since sept 20
cites consent and security
its own tools only
which door for your agent? πͺ
before: optimus watched us work
now: up to 20,000 bots learn by doing
one academy, every task practiced
every try feeds the model
fremont's first class is built
cars learned from millions of miles
optimus learns from millions of tries
same playbook, new students π€
one DNS misconfiguration
+ one model that found it
+ 11 days of paused training
+ red teams attacking their own walls
+ patches to kata, gvisor, docker
+ a cached web to practice on
= frontier training, no live internet, for now
trust a model that found its own exits? π
the fine print: OpenAI's Β£100M is compute + model credits with Rosalind Biodefense access, Anthropic's up to Β£100M is credits + support. the lab runs as an independent non-profit.
source: https://t.co/XZ2979lupd
UK built a lab for AI vs biology π§¬
Β£100M: government cash
Β£100M: OpenAI credits
Β£100M: Anthropic credits
DeepMind + Basecamp join in
world's first lab of its kind
spot threats early, speed up vaccines
credits are not cash, mind you
is this AI safety getting real?