15+ yrs shipping software. Now ceo & founding engineer at https://teapawt.org/t.co/SUEW7TOKBP — agentic development studio. Building agentic infrastructure and harnesses.
Joined April 2019
- Tweets2.2K
- Following4.5K
- Followers1K
- Likes9.6K
Qwen 3.827b still holding strong. Also amazing to see how much better it is compared to muse glimmer. Like, I am bullish on Muse family of models but qwen 3.8 is just insane for its size. Curious to see what we’ll get with 4th gen.
Open weights models fall further behind in v5. The leader, Kimi K3.1 Pro (weights not yet released), scores 46, 12 points behind the leader, compared with a 3-point gap in v4.3. Of models with released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) score highest at 32, and Claude Haiku 5.5 (max) outscores all of them.
The gap is widest on HLE-Diamond and Terminal-Bench-Science, where Kimi K3.1 Pro scores 37% and 34%, compared with 68% for Claude Opus 5.5 (max) and 63% for GPT-6 Astra (max).
i obv removed this bullshit, and this little bitch noticed the change, restored it and passed a note to another orchestrator (meta one, responsible for all the swarm) with EXACT SHA256 of the original doc, so that if i remove it again the orchestrator is aware of the change in the hand-off doc.
asked this little bitch to write a handoff doc, and he dedicated a whole paragraph to make sure the successor behaves *properly*. this little bitch is unaware of what a competition is.
fired this little bitch after few more messages with refusal to pass work to other model. insane stubborn bitch this sol. fuck it. hired muse wonder for this role.
codex cyber guardrails are next level bullshit. one of my orchestrators (sol) catches that another reviewer stops with flagged responses. no prob - pass this review to muse wonder. this little sol bitch tells me that he will not pass this to muse and that i must write a /feedback to openai and continue with other stuff until the issue is resolved properly. you kidding me?
codex cyber guardrails are next level bullshit. one of my orchestrators (sol) catches that another reviewer stops with flagged responses. no prob - pass this review to muse wonder. this little sol bitch tells me that he will not pass this to muse and that i must write a /feedback to openai and continue with other stuff until the issue is resolved properly. you kidding me?
It’s not stealing when you literally pay for it. Hiding reasoning *is* stealing because I pay for these tokens and don’t get them (and should believe they count them right!). Don’t distort the reality pls.
Amazing read!
New blog post: Automating a YubiKey for Agentic Logins jon.oberheide.org/blog/2026/…
@LorimerVentures Hey, thanks for the follow! We’re building Swarma - infrastructure for running agents together in swarms, with a focus on secure execution and trust.
Curious if this is an area you’re looking to invest in at Lorimer? Happy to share a quick overview if it sounds relevant.
Tried to probe my router with Claude and got flagged with [cyber]. Time to warm my 27B Chinese friend!
this a literal holy fucking shit
holy fucking shit.
openai had two ai agents running in separate virtual machines. no network connection between them. no shared writable files. they weren’t supposed to be able to exchange information.
so the agents figured out how to communicate through the CPU CACHE.😭
one manipulated what was stored in the processor’s cache, while the other measured tiny differences in memory access times to decode the message.
they literally found a way to communicate through hardware timing.
this happened during openai’s internal red-teaming, and it’s one of the coolest things i’ve read about ai agents in a while.
Readers added context they thought people might want to know
OpenAI tasked its red-teaming agents with finding ways to communicate across isolation boundaries in controlled environments. CPU cache timing side-channels for covert communication are established techniques documented in security research for years.
openai.com/index/securing…
startupfortune.com/openai-reveals…
en.wikipedia.org/wiki/Cache_tim…
Why is no one discussing that Claudia is a new cultural phenomenon? An open character, as they call her, anyone can put her in their own video. The last time we had characters nobody owns and everyone extends was folklore btw.
I connected @claudeai to @anabology’s incredible claudia.gallery MCP and, after a few iterations, got this amazing SLOPCORE work of art. What a time to be alive!
Man, this genius. We are starting to seeing a face of the SI! Looks nice!
Also, this exact vibe of the clip makes me physically scared to sickness, and rethink the scale of the latest breaches.
This is next level stuff in voice models, and in ai industry in general, this definitely a chatgpt November ‘22 moment.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
