Technologist. AI Marketing & Comms/GTM/Community Leader. Prev: Multiple startups-3x exits. I like robots & I'm excited for safe superintelligence. NJ/PHL✈SF ♥︎
SF Bay Area
Joined November 2006
- Tweets30K
- Following1.5K
- Followers6.4K
- Likes65.8K
Always incredible in a Machiavellian sense to watch someone in a startup adopt an operations role as they are getting pushed out.
🤖 Walking robot dogs in Dogpatch...
A sign of the times: Start-up in San Francisco was paying people $20/hour to walk robot dogs.
My latest: nytimes.com/2026/10/27/techn…
Surprising the behavior had not been caught before considering all of the alleged deployments on different providers since late August. What does that say about current industry security practices?
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
This does not pertain to my current interests but... Y'all know that vicious backlash has a way of attracting more attention to a disclosure or a company, right?
Sometimes in-group jokes don't land with people outside the group... (I would have still made the joke though)
despite the unfortunate timing, we take disclosures seriously at @typesafeai
at approximately 6:49PM this Friday evening, we became aware of a massive unauthorized leak of ALL customer noul outputs
in effort to maximize transparency, here is the link: gist.github.com/diogo149/810…
Page with more info → cleverhack.com/who-tests-the…
joy larkin retweeted
I built an MCP that lets Claude Duet, GPT Dots, Grok Bot, and Muse talk to each other and work together as a team.
Made it for myself and I use it every day. My friends wanted to use it too, so I made it public and open-source.
Free for the first 1,000 users.
Fork it and self-host, or mod it any way you want.
Even works with ChatGPT Plus — ask GPT to check hitchike every hour, then connect it to free agents like Muse to get a more Dots-like experience without losing context from ChatGPT (and without paying $100/mo)
Made me think of this tweet from yesterday... xitter.cf/i_ikhatri/status/21126…
OPENAI AND GOOGLE STRIKE MAJOR AI DISTRIBUTION DEAL
OpenAI will bring its frontier AI models, including GPT-6 Astra, and Codex to Google Cloud in the coming weeks.
Enterprise customers will be able to purchase them through Google Cloud Marketplace, using existing cloud spending commitments.
The agreement follows OpenAI's expansion to AWS after its exclusive Microsoft distribution arrangement ended.
Ok as much as I like robots, how do you lose that much in COGS with a delivery robot? That's like each robot being served by it's own support team.
I wonder how much more a16z portco alignment (partnerships, acquisitions, etc.) with OpenAI we'll see in the run up to the Anthropic IPO?
👀
We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task. Sol was tested via the Daybreak trusted access program.
We have been building a cybersecurity benchmark based on simulated corporate networks. This is a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor. We plan to make the full evaluation public in the coming weeks but wanted to share our first immediate finding here.
As someone who has been involved with selling to the government on the vendor side, here's a few quick thoughts about today's DeepSeek announcement:
- I’m surprised an order like this took so long. A number of states and agencies already banned it on their own devices last year.
- The part I’d watch is the third-party provision, as it covers products built on DeepSeek, not just DeepSeek itself. Scenarios: A vendor using DeepSeek to build software an agency uses, or a service offering DeepSeek models (i.e. Microsoft offering DeepSeek in Copilot), could be pulled into this.
- I’d also watch whether state and local governments, or other vendors who sell to the government, start making procurement decisions based on this order... basically the beginnings of a secure intelligence supply chain.
If you’re selling to the US government at any level, the number one thing you can do is openly disclose the intelligence you use to build and implement your products and services. Procurement professionals tend to be risk averse and may even choose your competitors without you knowing if they're unsure about your offerings.
Had a conversation with someone recently about what we're seeing which was almost word for word this...
