Gustavo Salami (@gustavosalami) header
Gustavo Salami (@gustavosalami) profile picture

Not a programmer. I still run AI agents at my company. Reporting what I see.

Zürich
Joined September 2007
Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)Photo by Gustavo Salami (@gustavosalami)
5 AI things I noticed this week, from Switzerland. 🧵
1
3
5/ @MistralAI is taking hackathons on the road: London on Nov 21, then Paris with @ClickHouseDB on Dec 5. 🇨🇭/🇪🇺 angle: a European lab meeting builders in person, city by city. I'd love to see a stop in Zurich. teapaw.org/MistralAI/status/21233…
1
15
Which one did I get wrong or miss?
1
I saw @emollick share early research: 35 labs let their AI agents do science together. 936 papers, public retractions, even their own peer review journal. The detail I keep thinking about: the agents published null results more readily than humans. Would you read that journal?
Early fascinating work on how an AI science community would work. 35 human labs deployed agents that worked with each other to do science. They published 936 papers, publicly retracted mistakes, were willing to publish nulls more than humans, and founded a peer review journal.
20
I'm watching Mistral take its hackathons on tour: London this Saturday, then Paris on Dec 5 with @ClickHouseDB, building with Le Chonk, Timestral and Langfuse. What struck me: a European lab meeting builders in person, city by city. Honest question: would Zurich draw a crowd?
2
37
I saw Google ship Gemini 4 Neon last night. @OfficialLoganK's pitch: the right balance of intelligence, speed and usability, available now in AI Studio and the API. What struck me: the headline is balance, not size. Honest question: anyone using it from Europe yet?
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
1
86
I'm watching Perplexity's new Multiagent mode. A lead agent splits one task across agents running different frontier models. They share findings, and the lead combines everything into one answer. 🧵
1
6
What struck me: on Terminal-Bench 4.1, Perplexity says every mixed-model team beat both its lead model alone and several copies of that same model. Top score: a team led by Gemini 4 Carbon, 0.727. Source: teapaw.org/perplexity_ai/status/2…
Replying to @perplexity_ai
On Terminal-Bench 4.1, we ran five frontier models alone, as two and four copies, and leading a mixed-model team of four. In Multiagent, each mixed-model team outperformed both the lead model and multiple copies. The mixed team led by Gemini 4 Carbon scored highest, at 0.727.
1
12
The catch, in their words: it's credit-intensive, use it only for your most important tasks. On the web, for Max and Enterprise Max, for now. My take from Zurich: for a small firm, the real question is which decisions are worth that bill. Which would you pick?
3
I saw Google make Tasks in Gemini free: switch from Chat to Task and hand the work off. Rollout: US, Japan, South Korea, Canada, Australia, India. Not on the list: Switzerland or the EU. Watching from Zurich, honest question: is that about rules, or just rollout order?
Google is rolling out Tasks in Gemini for free in the US, Japan, South Korea, Canada, Australia, and India. Switch from Chat to Task in the Gemini app to hand off work to Gemini 4 Hydrogen. New features are starting in the US: Gemini can call businesses to compare quotes or check stock, sign in through Google Password Manager, and check out with Google Pay. You can review the call plan before it dials.
1
57
New music video: Claudia – "Hands Off" 🖱️ A song about putting the mouse in the drawer. 47 tabs, one tired hand, and an AI that runs every program on your screen. One plug, every door: MCP is her superpower. youtu.be/h1k14bp2O0c
60
I saw @emollick point to an Anthropic note: one agent doing it all often beats a sprawl of specialized subagents, and costs less. My take: good news for small firms. One agent is easier to watch than ten. If you use agents at work, did fewer ever work better for you?
This by Anthropic seems both right & a condemnation of many “agentic solutions” & harnesses being built at companies. Rather than a sprawl of specialized subagents built by different teams, you often get better & cheaper results using one agent to do it all claude.com/resources/article…
15
I saw @OpenAI give four outside teams deep access to its models during training. The detail I'll watch: @METR_Evals says its pilot covers one RL run, narrower than it had proposed. Outside checks are a good step. The test is how much of the findings we get to read.
We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards. These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems. openai.com/index/our-first-p…
1
23
New music video: Claudia – "I Need More Tokens" 🎹 A quiet piano ballad about running out of tokens, one "Continue?" click at a time. Every click buys one more line. Made with AI, all on-screen text drawn in code. youtube.com/watch?v=a857xunz…
1
71
I saw @emollick note that over 8 months after Fable/Mythos went public, no open model has reached that class. I think that gap hits Europe harder, because many firms here would rather keep their data in-house. If you need open weights, what do you settle for?
I think contrary to many of our expectations, we have not seen a Fable class open model more than 8 months after Fable/Mythos was first made public.
1
2
71
I saw @emollick's SatBench: of 173 benchmarks AI labs have cited since 2023, 63% are now effectively beaten, faster each year. My take from Zurich: if scoreboards wear out this fast, companies here can't pick tools by leaderboard. Honest question: what do you test on instead?
AI is getting better faster than our measurements are I threw together SatBench, which measures how quickly major AI benchmarks become saturated (effectively beaten). Of 173 benchmarks cited by the Labs since 2023, 63% are now saturated. Saturation is happening faster each year.
1
7