Project Vend, Doppl and the Path to Medical Super Intelligence – Live and Learn #72

✨ Quote ✨

Friction, once an indispensable process of building competency and perspective, is vanishing to AI. A single prompt generates polished landing pages, completes brand systems, and ships working applications. The fruit that once required long, arduous labor now drops directly into our laps, ripe yet uncannily hollow.

– Willem van Lancker - (source)

🖇️ Links 🖇️

Project Vend by Anthropic. This is a write-up on an experiment Anthropic ran in-house with one of their Claude models. They tried answering the question: Can Claude run a small business? So they gave Claude access to order items and set the pricing for a small vending machine, and then let Claude loose to run all the operations. The results were... interesting. In the end, the model went bankrupt, making all sorts of weird decisions. This highlights that models are not quite there yet and at least for now, can't be used as middle managers in real-world scenarios. To me, the most interesting and funny fact from the report was, that at some point, Claude started hallucinating—thinking that it was a real person. It even started writing Slack messages to Anthropic employees telling them that it is by the water cooler wearing a certain outfit—waiting for them to pick up their order "in person". Note that all the restocking, supplying, and actual warehousing was still performed by real humans in this experiment. Amazon just announced that they employ more than 1 million robots, so maybe fully stocked automatic stores will be a thing of the future.

Doppl by Google Labs. This experimental app by Google lets you upload a photo of yourself and of pieces of clothing that you like. And then the AI model will try to match how this particular piece of clothing will look on you in real life. It's an interesting direction, and I wonder where and how they will integrate this into online shopping experiences.

The Path to Medical Superintelligence by Microsoft. Microsoft is thinking about how to use LLMs and algorithms to help doctors do better work and make more informed decisions. Their research, up to date, highlights that AI can help, because it can provide a lot of context and specialized knowledge quickly—much more than any single doctor with only one specialization could know. They also say that this doesn't mean AI is going to replace doctors any time soon, but that it can enhance their decision making. But looking at this robot surgeon doing a first real-world operation, autonomously, makes me think twice about that.

Gemma 3n by Google. This newest model in the open source Gemma series is targeted to be small and performant and run fully on device. What is insane to me is that it still has audio and video ingestion capabilities (check out their demo video) while maintaining really, really good performance, for its ultra small size and memory footprint. On-device AI will be a thing... a super powerful thing, transforming our smartphones into even crazier powerhouses than they already are.

Inference-Time Scaling and Collective Intelligence for Frontier AI by Sakana. All the research coming out of SakanaAI these days seems to be, if not immediately useful, at least highly interesting. In this paper, they created a way for different LLMs to cooperate together. This adds a third dimension to inference time scaling (besides number of trials and time spent "thinking")—namely, how many models you let cooperate. Letting multiple models tackle the question together with their AB-MCTS architecture improves the performance. They got a ARC-AGI 2 ~30% pass rate, compared to ~26% from only querying o4-mini over and over again. Notably, this pass rate is for 250 submissions to the questions, a big difference from the standard ARC-AGI 2 guidelines, which normally only allow 2 submissions per question. In other words, this approach doesn't solve AGI yet, nor does it saturate the ARC-AGI 2 benchmark. But it is still interesting that something akin to human collaboration (different "minds" thinking together) is even possible in the first place with AI.

Neuralink Summer 2025 Update by Neuralink. The team assembled by Elon Musk has been cooking, and they now have had 5 successful human transplants and are planning to rapidly scale up both the number of participants in their clinical studies as well as the scope of their trials. Their ambitions are set on deeper brain regions, and the amount of reading and writing they can do to neurons is about to increase as well. All while they upgrade their robot installation hardware to make the progress of implanting the devices as seamless as possible. What they showcased looks so much like crazy science fiction that it seems almost unreal, but it makes me excited and rather a bit anxious for the future. Brain chips that can literally read and write signals to the brain are going to be a thing, kinda soon. On this note, I feel like it is time to re-read the 2017 Neuralink post by WaitButWhy.

🌌 Travel 🌌

The last two weeks killed my spirit because the TEMBR had sections with even more insane headwinds, a good amount of rain, cold, and landslides, too. All this got me to the point where I stood on top of a mountain, rain hitting my face, screaming against the wind, and hating the world. In the end, I gave up and just took a bus to reach Cuenca, where I could leave my bicycle to return, maybe the next year. The views, especially up around the Chimborazo volcano, were worth it in the end, though. Even cycling to the refugio at 4850 meters elevation, gasping for air, and almost freezing my fingers off in the morning—all worth it. Those are memories I wouldn't want to miss. I'm happy to be returning home and seeing all my friends again, though, but at the same time, sad because there is so much of South America left to discover. But that's for another time.

🎶 Song 🎶

In the Real Early Morning by Jacob Collier

Youtube Music | Spotify

---

That's all for this time. I hope you found this newsletter useful, beautiful, or even both!

Have ideas for improving it? As always please let me know.

Cheers,

– Rico