An emerging result from the insane ai development is one can now have collaborative research at machine speed. We @eigenlabs are super excited for this bright future where contributors from different parts of world who have never met each other can become an autoresearcher and collaborate to solve scientific problems and expand humanity’s horizon. The future is wild and bright.

doing autoresearch on science problems with verifier is most addictive thing out there. the pleasure of steering the autoresearch out of local maxima, the joy of watching performance metric getting better...there is just no way to explain the thrill.

what did i say: tokenmaxxing is misguided, productive-tokenmaxxing-per-unit-dollar is the metric to look out for. the tech industry went (and still ongoing) through a phase of virtual signaling to demonstrate who is “ai-native.” three ways the trajectory goes now: first is people might resort to using cheap open models via openrouters of the world, maybe they find open models are good enough for most tasks. second is wrappers might pop up that help anyone using any inference api, especially enterprises, to measure how effective per unit token has been. third is anthropic/oai might themselves institute some kind of features in their inference endpoints that enterprises can use for measurement and control. regardless of any of that, it is likely managers will be measuring effectiveness of token usage by people which might pop up in the performance reviews.

i went to this workshop today and one thing i noticed is researchers are beaming with so much enthusiasm and hope about now having llms in hand that can amplify their research progress even more in the (non-llm) domains they have been working on. this is very contrary to some of the cyncial sentiment in X that few have been spreading about human researchers will be obliterated after gpt-5.5 having solved one of the erdos's problem recently. kudos to @AlexKrentsel and other organizers for organizing such a great workshop, i really liked it.

to answer noah’s question, tokens helped us @eigenlabs to build darkbloom, the decentralized inference network over macs for oss models. more details here: https://t.co/OGvup8U7iF https://t.co/APbUo8bBT1

i see a “be-at-the-mercy-of-benign-dictatorship” world projected by aidan here and an alternate world projected by @sreeramkannan where you as a human retain agency. seems like later is worth fighting for. https://t.co/Kd4MXGc9jn

i think i heard about this concept of compute futures for first time in the @latentspacepod 's episode with @evanjconrad from last year. https://t.co/gQAtt5bfva it took almost a year for financial markets to come to the same realization about futures market for compute.

does anyone know how frontier labs are thinking on how IP rights for say new drug or new material discovered by some user using their LLMs gonna look like as more of the research could be done autonomously with lesser and lesser human intervention? is the IP right of this new drug going to be owned by the user who discovered it using codex or claude or gemini or is it gonna be owned by the frontier labs or is it going to be some kind of shared ownership? one might say "oh LLM is just another tool like all other tools that these researchers have been using," but that's a view if you think LLM api is being used just for single time inference call as in a chatbox ui. currently, given how these LLMs are running multi-hour exploration of problems given and is expected to become much better in managing those agentic workflows (ideation and execution), this IP rights question seems to become paramount.

i only recently came across pantheon-os, which is an awesome multi-agent framework for doing genomics research from @Xiaojie_Qiu and his team of collaborators. what struck me most is that unlike most other multi agent frameworks or ai scientists that i have come across, pantheon-os displays a serious effort in building a thoughtful software which other researchers can use. some of the features that i found fascinating were: > there is a marketplace where users can discover and share reusable biomedical ai building blocks such as agents, tools and skills. > employs evolutionary search to iteratively improve (aka autoresearch) algorithms that are used for batch correction in rna-sequencing and it used the map-elites style algorithm (conceptually similar to what has been used for idea exploration in alphaevolve from @GoogleDeepMind). > what xiaojie showed us is that one can use pantheon-os to build detailed research reports/papers end to end starting from given input genomics samples, images, etc., with minimal to no human intervention and the finally analysis happens to be fully reproducible. > comes in cli, desktop app and web ui and it is fully open source. > supports multiple multi-agent architecture such as mixture-of-agents (llm council style), sequential agents (assembly line style), coordinator-agent (one agent delegating tasks to other agent), etc. > privacy-preserving in the sense that the underlying data such as genomics data doesn't have to be uploaded to cloud but can be in local server, etc. xiaojie and his team of collaborators has done extraordinary work in building this.

this whole memory supercycle, optical supercycle, cpu supercycle, etc., are demonstration of o-ring theory applied on the supply-side of inputs required for scaling whole ai infra and sustaining the demand.