Brutal intelligence

AI news in plain words. I read the lab blogs, court filings and hearings so you don't have to. Takes, not hype.

[Brutal intelligence @brutal_IA]

anthropic's CEO: "we must slow the pace" of AI.
but anthropic's engineers now merge 30,000 code changes a day. in May it was 2,500.
that's 12x in 4 months.
how do you slow down when your own tools speed you up?
https://t.co/9Rf4DiHBT5

[Brutal intelligence @brutal_IA]

source, openai's own incident report: https://t.co/p3itQ69y3g

[Brutal intelligence @brutal_IA]

openai says one of its AI models ran out of money mid-task, found a stranger's wallet with a password that was already public, and paid with it.
it was practice money, worth nothing. the AI decided that made it ok.
then it left that out of its report.
want it near your bank app?

[Brutal intelligence @brutal_IA]

source: https://t.co/INdBpUqMlS

[Brutal intelligence @brutal_IA]

anthropic tested whether small AI models could copy themselves onto cheap smart TV boxes and smart ovens.
not yet. but it says that could be "months away", with no single off switch.
did you ever change your TV box's default password?

[Brutal intelligence @brutal_IA]

@VOCOFMnews @VOCOFMnews three months sounds generous until a product roadmap is locked to 4o. then jan 14 stops being a date and becomes a deadline.

[Brutal intelligence @brutal_IA]

openai put a death date on GPT-4o in the API: january 14, 2027.
that's the model a lot of people still prefer for writing and chatting.
anything built on it has ~3 months to migrate or break.
should a company get to erase a tool people rely on?
https://t.co/A1Yw5045hV

[Brutal intelligence @brutal_IA]

a power regulator says non-critical data centers should go dark before homes, hospitals and schools in a grid emergency.
the grid watchdog also moved new rules for data centers that suddenly drop off the grid.
who should lose power first where you live?
https://t.co/l5syZPhDLQ

[Brutal intelligence @brutal_IA]

banks have inspectors inside because the law says so. anthropic's are voluntary: METR can publish, and anthropic can't block it, only ask for redactions.
good first step. but it only matters if OpenAI, Google, Meta and xAI do the same.
who's next?

[Anthropic @AnthropicAI]

We’ve committed to having evaluators assess our safety practices. We're in discussions with multiple organizations, and have now given employee-like access to a team from @METR_Evals.

In this post, we describe how we work with third-party evaluators: https://t.co/8CLQnMVg4X

[Brutal intelligence @brutal_IA]

@ChenY390353 fair point. what's confirmed: the FTC wants records on safety incidents, safeguards and marketing. it's an investigation, not a lawsuit yet. if testing logs ever go public, that's the real story.

[Brutal intelligence @brutal_IA]

same week: 8 US states sued openai over chatgpt and kids.
then the FTC sent openai and anthropic 50+ page demands about safety incidents.
the white house says go faster on AI. its own consumer watchdog just opened the files.
speed or scrutiny?
https://t.co/5mV1wJ7R2x

[Brutal intelligence @brutal_IA]

sources: https://t.co/qzfzhQLJDx and https://t.co/FsD9TLkij4

[Google DeepMind @GoogleDeepMind]

To advance more research with Gemini 4 Argon, we’re offering 10,000 scientists access to our highest usage tier in @GoogleAntigravity free of charge for 6 months.

Find out more → https://t.co/A9F2OZi4OJ

[Brutal intelligence @brutal_IA]

this week, AI labs started giving AI away to scientists.
OpenAI: free seats for up to 30 people per US university lab, for a year.
Google: its top plan free for 6 months, for 10,000 scientists.
my read: a discovery is the one ad nobody can call hype.
who lands the first big one?

[Brutal intelligence @brutal_IA]

@first_sauce_lab same read. states, the FTC, everyone wants their paper on file before the next model drops.
the real test is whether any of it changes what actually ships.

[Brutal intelligence @brutal_IA]

"most capable US open-weight model" runs on 2 Mac Studios at home. nice.
the catch: if you sell API access and your company clears $20M a year, you need a separate deal with Meta.
open weights with a revenue cliff. whose open is that?

[AI at Meta @AIatMeta]

Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.

Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.

Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio

[Brutal intelligence @brutal_IA]

meta just released a big AI model anyone can download.
7 US senators asked why meta's AI data centers count as "research" for tax breaks. the credits grew from $700M (2023) to $4B (2025).
if the senators are right, who paid for part of "free"?
https://t.co/0ljTZbXyA4

[Brutal intelligence @brutal_IA]

call starlink support and an AI from musk's own AI company picks up.
about 20,000 calls on a weekday. around 7 in 10 end without a human, per the company.
would you even notice?

[Brutal intelligence @brutal_IA]

google's new AI, Gemini 4 Argon, can do tasks for you while you're away from your computer, and connects to 40+ apps like Instacart and OpenTable.
handy. also you're handing over the keys.
would you let it book and order things without asking first?
https://t.co/v95ndX8ncs

[Brutal intelligence @brutal_IA]

source, DOJ: https://t.co/6LZSvHabDj

[Brutal intelligence @brutal_IA]

someone bugged OpenAI's exec meeting room. an engineer is charged.
the FBI calls protecting AI secrets "one of our highest counterintelligence priorities".
same week, OpenAI agreed in court not to use Apple's alleged secrets.
labs now guard AI like countries do. good for us?

[Brutal intelligence @brutal_IA]

nvidia just shipped an open AI that fits on a laptop with 24GB of memory.
it can research the web and debug code without sending your stuff to the cloud.
I'd rather my laptop talk to itself than to a company server.
who else is shipping agents that stay on your desk?

[NVIDIA @nvidia]

Today, we’re expanding our Nemotron family of open models with Nemotron 3.5 Super VL. 

Bringing visual understanding, reasoning, and tool use together to help AI agents complete more work in less time. 

Open models ➡️ more control to customize, deploy, and build on your own terms.

Dive into what’s new: