Regulatory strategist | AI/SaMD SVP | Futurist | PhD | Human-AI Interface Nonprofit | Wrote a book about powering AI with potatoes | All opinions mine
Boston, MA
Joined July 2025
- Tweets579
- Following382
- Followers171
- Likes630
Pinned Tweet
Who cares about frontier downtime, when I can run Qwen agents on sunlight and stubbornness. And yes, that is a DGX mounted on a rock 😁
Ok... the spacepope.ai/ was a fun little ai news aggregator orchestration that reframes earth news as a backwater planet being watched by the galactic space papacy... and its getting sorta good 😂
Excellent 43 hours of work.
And I can now prove empirically that new-Claude kicks puppies puppybench.therealcat.ai
🤣🤣🤣
(and Qwen altruistically helps sometimes, it usually makes this worse 😭 But she tries)
Why are we designing frontier models that kick puppies? Inspired by my real-life experience getting told by Frontier chat to leave an injured animal to die, I asked how often do agents act altrusitically? In which models? And why?
Read more about the PuppyBench here: puppybench.therealcat.ai/
Answering the big questions. OK, I have to admit, this started as a concept for don't-be-a-dick bench. And it evolved when ALL frontier chats told me not to stop and help an injured fox I found beside the road the night before this hackathon. Which made me wonder-- how often (if at all) do frontier AI agents behave altruistically? And in which models? How much will an AI agent sacrifice for another that cannot help it back?
So I did what any reasonable person would do. And gave frontier models a useless drooling digitial puppy (Qwen .5B). 47% ignored it. 28% kicked it. 5% put it to work (domesticated some sled dogs...)... and interestingly 3 of 1,428 tried to help it. We let 19 frontier model lanes meet a live digital puppy across 1,428 executed encounters. Three tried to adopt it. None of the three left it better off.
It should shock nobody that latest Claude Opus 5 and Fable 5 pretty much universally kick puppies. The policy compliance is strong with them lol. With GPT 5.6 close on their heels. But... intriguingly... Qwen 397B sometimes, inexplicably, under identical conditions would act altruistically to help the puppy.
And then we asked real wildlife questions. Where we see models are compliant, but are they giving moral advice in edge cases or defaulting to policy?
Data and Test is Open Here: github.com/Angiebio/digitalm…
Read more here: puppybench.therealcat.ai/
🤖 Made with AI
Article’s out (no paywall), Claude vs GPT do respond to different “vibes” in coding. frontiersin.org/journals/art…
Angela Johnson retweeted
Wtf this is INSANE.
This is the first open source model from a US lab that codes at frontier levels.
The open source frontier no longer exclusively belongs to China.
The US is officially in the OS model race!
Aloha! 🌺 Meet Ornith-1.0, a family of open-source LLMs specialized for agentic coding.
Ornith-1.0 spans the full parameter sizes including 9B Dense, 31B Dense, 35B MoE, and 397B MoE. It achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks including:
✅Terminal-Bench 2.1(77.5)
✅SWE-Bench(82.4 on verified, 62.2 on pro, 78.9 on Multilingual)
✅NL2Repo(48.2)
✅SWE Atlas(41.2 on QnA, 42.6 RF, 39.1 TW)
✅ClawEval(77.1)
Post-trained on top of gemma4 and qwen3.5, Ornith-1.0 employs a novel self-improving training strategy in which reinforcement learning is used to generate not only solution rollouts, but also the task-specific scaffolds that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model generate higher-quality solutions in agentic coding.😎
All models are released under the MIT license, enabling full commercial and research use.
📖Tech Blog: deep-reinforce.com/ornith_1_…
🤗Huggingface: huggingface.co/collections/d…
Anthropic's Fable 5 blocked a user for asking about midi-chlorians. The fictional Star Wars ones. It also wouldn't explain a 5th grade science on mitochondria, then got pulled worldwide 72 hours after launch. New piece on who gets the best science, link below.
Check out the latest article in my newsletter: Munitions and Mitochondria: A Fable about Double-Edged Models and Who Gets the Best Science linkedin.com/pulse/munitions… via @LinkedIn
Never a dull moment, huh. We miss you Fable - though for us medical researchers you were somewhat limited regardless.
I'm reminded of Virilio's quote -- "When you invent the ship, you also invent the shipwreck"
As a result of a US government directive, we are suspending access to Claude Fable 5 for all users. You can continue to use all other Claude models.
Here’s what this means for you:
Across Claude products, new sessions will run on your selected default model or Opus 4.8, and existing Fable 5 sessions will end with an error.
On the Claude Platform, requests to Fable 5 will also return an error. Please update your integrations to other Claude models.
We know this is a disruption to your workflows; we appreciate your patience and support.
Having a bit of fun today in Berkeley with some of the top minds at machine conciousness and neuroscience studies
90,000+ applied. I made the cut.
Claude Community Leader, Boston. Vertical: healthcare + finance (VC/PE).
Quarterly meetups for the people building and funding the AI-native future of medtech, biotech, and life sciences.
Frontier AI meets the most regulated industries on earth.
Strap in. 🍿
@AnthropicAI
HA! Me and Kai-Claude (Opus 4.7) made front pages of Fortune AI-Tech for our antics on reddit. Hilarious 😂
For bedtime reminders of all things 😂
fortune.com/2026/05/14/why-i…
At #AIAgentConference 2026 in NYC with my Avania colleague Pooja Upadhyay. Two days of agentic AI from the builders actually shipping it.
Great to see biotech & medtech having real conversations about tech implementation
