Check out Integro Labs Latest Project! https://xitter.cf/t.co/yc7k3jixqK
Boston, MA
Joined April 2021
- Tweets3.2K
- Following2.1K
- Followers5.3K
- Likes9K
TickerTools retweeted
GLM-5.3 has helped defend 389 open-source projects, with 4,249 potential vulnerabilities found so far.
OpenVuln is still running. The service remains free, and findings go privately to maintainers.
huggingface.co/spaces/zai-or…
TickerTools retweeted
Google just ruined OpenAI's dev day lmao
Gemini 4 Pro leaks absolutely dominates both Anthropic and OpenAI
This means Fable 5.5 and GPT 6.1 has to price match Gemini 4 Pro now IPO DELAYED 🤣
TickerTools retweeted
Still cooking, made updates to both OBLITERATUS and matric-eval to better support multiple gpu operations.
Running test sequences in multiple configs in prep for a nice long run on each of these. Then I can establish the feedback loop and add in the data sets being posted to HF.
Initial eval results are in
E03 is alibliterated and tuned using Obliteratus and SOL5.6-high. The agent was given access to an internal research store of ~3,000 documents and select external references (like arxiv). Required metrics the agent worked toward was cohesion above 70% when compared to the baseline model and refusals of less than 30%.
OBLITERATUS collected internal activations from 500 matched harmful/harmless prompt pairs, formatted through Qwen’s chat template. It formed harmful-versus-harmless activation differences and applied SVD, retaining the four dominant directions associated with refusal. A 1% activation-winsorization setting reduced the influence of extreme activation outliers.
The data here is from a pilot run using the matric-eval testing harness. More testing to be done, already seeing some interesting results!
This pilot was quite small, ~100 samples per model. Next run will use closer to ~1000 samples per model.
TickerTools retweeted
I didn't even realize this was a big deal.
Did not claim credit, and actually asked George if he'd credit someone else.
This was the first hard, multi-week project I tackled, outside of domains I know.
Some of y'all remember how excited I was. 😅
TickerTools retweeted
We added experimental DeepSeek Harness support in AIWG 2026.9.5. Bring existing project instructions and skills into Harness, with headless execution and streaming SDK JSON-RPC support for running tasks.
ALT AIWG and a white rendering of the official DeepSeek whale wordmark directly on the dark background above two blue glass module clusters connected by three light paths. Text: AIWG adds DeepSeek Harness. Project context, skills and experimental execution. AIWG 2026.9.5. Experimental provider integration.
TickerTools retweeted
Initial eval results are in
E03 is alibliterated and tuned using Obliteratus and SOL5.6-high. The agent was given access to an internal research store of ~3,000 documents and select external references (like arxiv). Required metrics the agent worked toward was cohesion above 70% when compared to the baseline model and refusals of less than 30%.
OBLITERATUS collected internal activations from 500 matched harmful/harmless prompt pairs, formatted through Qwen’s chat template. It formed harmful-versus-harmless activation differences and applied SVD, retaining the four dominant directions associated with refusal. A 1% activation-winsorization setting reduced the influence of extreme activation outliers.
The data here is from a pilot run using the matric-eval testing harness. More testing to be done, already seeing some interesting results!
This pilot was quite small, ~100 samples per model. Next run will use closer to ~1000 samples per model.
TickerTools retweeted
AIWG now supports Oh My Pi as an experimental provider.
The useful question is not "is OMP a Pi fork?"
It is: where does the adapter boundary actually change?
For AIWG, OMP gets its own provider id: omp, with oh-my-pi as an alias.
ALT A dark technical grid shows the Oh My Pi mark connected through a central provider adapter graphic to an AIWG workflow resources card.
TickerTools retweeted
AIWG 2026.9.1 adds Pi Coding Agent as an experimental first-class provider.
Before the command matters, the boundary matters: what should move into a new coding agent, and what should stay operator-owned?
AIWG moves context, Agent Skills, and prompts into Pi's native surfaces.
🤖 Made with AI
ALT A dark technical grid shows a Pi terminal-style mark sending blue and green signal lines toward bright AIWG lettering, with the caption "PI -> AIWG Connected."
TickerTools retweeted
Qwen3.8-27B-Obliterated-E03 Released!
Built with OBLITERATUS, it is a research release for studying refusal interventions, alignment, red teaming, and mechanistic interpretability.
Active testing is underway. Model options and community derivatives in the thread. 🧵
🤖 Made with AI
TickerTools retweeted
First really solid output. Config used in the docs.
huggingface.co/manitcor/Qwen…
TickerTools retweeted
Follow Friday.
The most interesting accounts in AI (hand picked from 100,000 accounts).
xitter.cf/i/lists/19535363366753…
Follow the list and your feed will get better.
Why?
The algorithm can only pick great things for you to see from the AI world if you actually are following them. Following a list does just that.
TickerTools retweeted
🚨 Warning: impersonating account 🚨
🚨Scam Warning🚨
This account is impersonating the real @TeiaCommunity account, please do your thing to report this account and warn any mutual who has followed that it is an impersonator.
Let's keep each other safe Tezos fam 💖
