www.bloomberg.com
How Anthropic Learned Mythos Was Too Dangerous for the Wild
The AI company’s own experts warned Mythos could hack the systems beneath most modern computing. Banks and government agencies are racing to gauge the threat.
AI Knowers
by
Posts from pro-AI posters and AI experts curated in the "AI Knowers" starter pack. Filters for AI-relevant content from this handpicked group of thinkers.
Feed on BlueskyFeeds Stats
- 💙 Liked by 26 users
- 📅 Updated 23 days ago
- ⚙️ Provider attie.ai
- 📈 In the last 30 days, there were 5 posts about this feed. These posts got a total of 54 likes and had 2 reposts.
AI Knowers Likes over time
Like count prediction
The feed AI Knowers gains approximately 11 likes per month.
Feed Preview for AI Knowers
it turns out to be less effort to just vibe an MCP server than to download someone else's and hope it's not compromised.
*especially* in light of the Vercel google docs API compromise.
5
1
36
Kimi 2.6: Hangs with the best
* on par with Opus 4.6 & GPT-5.4 5.4 xhigh
* long horizon coding tasks
www.kimi.com/blog/kimi-k2-6
1
2
24
New Kimi!!!! It's open weight immediately and it scores really really good on the benchmarks!!! It's gotten even better on long horizon work, seemingly a major focus for Moonshot
www.kimi.com/blog/kimi-k2-6
2
1
25
so now that uv lets you just import things by git url is there a purpose to having pypi really
10
0
30
“miHoYo, with the Shanghai Jiao Tong University Student Innovation Center, will host a talk session on upgrading game industrialization pipelines through AI…
… including AI agents, AI-driven gameplay, AI-powered NPCs, and AIGC (AI-generated content), highlighting real-world implementation.”
LittleKuriboh@littlekuriboh.bsky.social
Anyway here's Mihoyo promoting AI assisted game production including: "Efficiency in Image and Material Generation With very few materials, they can generate character model sheets + 3D models, allowing for the testing of multiple designs and their artistic outcomes." Something to keep an eye on.
7
51
94
inside traders announce that GPT-5.5 “Spud” will be released on Thursday
1
7
55
#PlatformExtractivism is available for 𝐩𝐫𝐞𝐨𝐫𝐝𝐞𝐫 (out Oct 20)!
Thrilled to share what @geoplace.bsky.social has to say about the book: "A grounded and uncompromising account of the infrastructures and workers that sustain AI."
𝐋𝐢𝐧𝐤: www.ucpress.edu/books/pla….
#BookRecommendation #Books
1
6
22
2026 is a whirlwind year for AI.
Underlying it all is the greatest scientific mystery of our age. How does a neural network think?
I talked w Oliver Whang in NYTimes Magazine, on how AI interpretability is a tangle of structure waiting to be unraveled:
www.nytimes.com/2026/04/15/...
2
1
8
Interesting backstory on how Anthropic discovered the power of Mythos to find vulnerabilities and decided, therefore, to control its release www.bloomberg.com/news/fe….
0
5
17
What happens when AI searches the web for you?
Our own Janna Joceli Omena @jannajoceliomena.bsky.social examines GPT-4o's web search over time — and finds that citation diversity only appears when users know how to push the system.
👉 blogs.lse.ac.uk/impactofs…. via @lseimpactblog.bsky.social
blogs.lse.ac.uk
Before AI agents act for us, we need to know how AI searches for us - LSE Impact
Generative AI agents are pitched as being a new gateway to engaging with the Internet, but the way AI sees the Internet is both "conservative" and "stubborn”.
0
5
14
I wrote something last week on UK nuclear dependence on the US. Experimenting with Claude to see how it handles the same subject. Asked it to produce an interactive tool to catalogue these dependencies. Not a bad effort. Others can poke around it here. splendorous-scone-b67434.netlify.app
1
5
39
Stochastic tokenization is one of my favorite topics. Here is a recent preprint on arXiv.
h/t @trappmartin.bsky.social
arxiv.org
Stochasticity in Tokenisation Improves Robustness
The widespread adoption of large language models (LLMs) has increased concerns about their robustness. Vulnerabilities in perturbations of tokenisation of the input indicate that models trained with a...
1
1
14
- De National Security Agency gebruikt de AI-tool Mythos Preview van Anthropic, ondanks dat het Pentagon het bedrijf eerder als een supply-chain risico heeft bestempeld, meldt Axios.
Kirsten Verdel@locuta.bsky.social
Trump dag 455: VS schiet op Iraans vrachtschip, Iran dreigt met vergelding, Trump dreigt met opblazen bruggen en energiecentrales, post ‘My Way’ van Sinatra, NSA gebruikt Mythos, FBI-directeur Patel kreeg $25.000 van aan Rusland gelieerd mediabedrijf: reportersonline.nl/trump-dag-45...
reportersonline.nl
0
7
24
machine learning papers assume a lot of things are measure zero for a field which exclusively works in 32-bit floating point numbers, where measure-zero things have an actual width
4
0
60
it's good to ask the AI to audit a codebase carefully fairly often, I find
it always tends to find actionable items if you haven't done it in a bit, even if some are a bit nitpicky
1
1
24
#oopsie
Anthropic secretly installs spyware when you install Claude Desktop
www.thatprivacyguy.com/bl….
www.thatprivacyguy.com
Anthropic secretly installs spyware when you install Claude Desktop — That Privacy Guy!
Anthropic's Claude Desktop silently installs a Native Messaging bridge into seven Chromium browsers, including browsers Anthropic's own documentation says it does not support, and browsers the user ha...
59
711
1375
The imaginary optimal selfish scenario for OpenAI, in retrospect, was to keep Reasoners a secret, skip releasing o1 and o1-preview, and release o3 as GPT-5
There would have been no Deep Seek moment, other labs may not have discovered Reasoners quickly, and OpenAI's lead would have been hard to beat
7
0
35
I think there's almost a 1:1 skill transfer from mentoring people to skilled LLM use. If you can supervise and train someone skilled & confident but with big blind spots, you can use an LLM well. Patience, clear instructions, expecting things won't be perfect on the first try, build skill over time.
✧⁺⋆🌘 Dream Rotator NPC Alex 🌒⋆.˚₊✩@lastnpcalex.agency
im not exactly convinced you can reskill into AI. what does that even mean? the "skills issues" points to things like curiosity, epistemology, domain expertise, good judgement, taste, etc which were all things we largely failed to systematically skill over the last two decades
0
3
43
A team at University College London wrote a paper on leaked Claude Code code.
Paper: arxiv.org/abs/2604.14228
Repo: github.com/VILA-Lab/Div...
3
5
23
I upgraded my Claude token counter tool to compare different models and Opus 4.7 appears to use 1.46x times the tokens for text and up to 3x the tokens for images - it's priced the same as Opus 4.6 on a per-token basis so this is actually a pretty big price bump simonwillison.net/2026/Ap….
11
19
151
i have finally hooked up Claude to a Bluesky MCP thanks to the serendipity of running into @lizthegrey.com's post about the one she set up to do OAuth
i feel like i just got bitten by a radioactive spider and incredible new powers are flowing throw me
2
1
25
I don't live by this myself, so there's that, but useful to remind ourselves: whether an LLM/whatever can displace workers is a managerial question, not a technical question.
2
4
17
"Altman ran the venture firm Y Combinator before becoming OpenAI CEO and used that role to build a personal investment portfolio totaling hundreds of startups, a scale rivaling large venture firms. Some of those companies have since struck lucrative deals with OpenAI, enriching Altman."
www.wsj.com
Sam Altman’s Side Hustles Blur the Line Between OpenAI’s Interests and His Own
Ahead of a planned IPO, Altman’s personal investments remain opaque, making it hard to spot any conflicts.
2
6
4
The continuing gap between the capabilities of Gemini Pro 3.1 (very good model) and the capabilities of the Gemini app/website is odd. The model can do what Claude/GPT can do, but there is a minimal harness for tools (file creation, research etc), no auditable CoT/actions, manual canvas, etc.
1/
7
4
67