We are building this at @aitomatic, alongside partners like @honeywell, @panasonic, @furuno_global, @VPBankVN, and @fpt_software. We call it the Cognitive Ontology. It is the knowledge graph for the agentic enterprise.

Today marks one year of @thealliance_ai! Over the last year, @aitomatic and our international community have been working to develop #Al collaboratively, transparently, and with a focus on safety, ethics, and the greater good. We're just getting started. thealliance.ai/blog/our-firs…

Meanwhile we will have to settle for DANA as our state-of-the-art emulation. arxiv.org/abs/2410.02823 tl;dr: DANA (Domain-Aware Neurosymbolic Agent) blends domain knowledge with a neurosymbolic framework beyond standard LLMs, achieving 90%+ accuracy and consistency in complex tasks, with strong potential for precision industries like finance and semiconductors.

People underappreciate what you mean by cat-telligence esp. for Planning & Reasoning, @ylecun. (Via tiktok.com/@crabfren/video/7…)

“Open foundation models like Granite, Llama, StarCoder, GHP-MOFassemble, ESM-Fold, and SemiKong demonstrate the transformative potential of open weights in accelerating research and driving innovation across various sectors.” Statement from #AIAlliance Regarding Commerce's Final Report on Open-Weights Models thealliance.ai/blog/dept-com…

Industry’s first-ever open-source Semiconductor domain-specific model “SemiKong”, being announced at #SEMICONWest by me (!) and colleagues from @aitomatic @TokyoElectron_J @fsoft_aicenter. This in unprecedented for the semiconductor industry which is infamous for competitive secrecy. Outweighed by the benefits of shared R&D at the foundational layers. While encouraging proprietary differentiation and competition at the company/tool/process/device-specific layers above. This industrial-scale collaboration led by @aitomatic is made possible by the #AIAlliance and its members. Quite significantly: thanks to open foundation models like Llama3 cc @finkd @ylecun @AIatMeta who is also part of the AI Alliance.

arxiv.org/abs/2404.11792 Real-Reasoning RAG, or Where to Get Performance Gains Out of RAG (from @aitomatic & @IBMResearch, #AIAlliance): • Fine-tuning the retriever model gives you more bang for your buck than the generator model. • Employing reasoning yields significant gains, esp. on compound/abstract queries. • Planning & reasoning is where everyone will have to go, even just to solve enterprise Q&A. cc @ylecun. Related, “Compound AI”, cc @matei_zaharia. HT @zooey_nguyen et al.

linkedin.com/pulse/agentic-a… Explore how Agentic AI is revolutionizing problem-solving across industries with this latest analysis. From semiconductors to energy, Agentic AI is a game-changer for leading CxOs seeking to advance their business strategies. Learn more about this transformative technology and its critical role in real digital transformation. #AgenticAI #BusinessStrategy #RealDigitalTransformation

Planning and Reasoning coming soon to a repo near you. github.com/aitomatic/openssa Cc @aitomatic

At Davos #AIHouse I spoke about: » The AI Alliance thealliance.ai and its mission to Build, Enable, and Advocate for open development, open science, and open AI. » That the best way to achieve Safety, Reliability, and Trust is through open source. » That it is erroneous to conflate open/closed with managed/unmanaged. Indeed, the best way to manage is through open development. » That the best way to manage AI risks is to look at two main elements: (1) the point of application/not technology, and (2) the incentives around their impact/not intent. » Likewise with ethical/unethical AI—community development tend to be held to higher ethical standards than behind closed doors. » That the industrial world has no illusions about the risks of putting probabilistic components into life-critical systems—control loops are in place for deterministic outcomes. » That it is by now well understood that LxM are Level-1 (inferencing) intelligence that require an external control loop to achieve Level-2 (reasoning) intelligence—and Level-2 systems will be far more deterministic. I knew about but could not reference #AlphaGeometry yet—it is a neuro-symbolic approach to the above Level-2 intelligence. To @GaryMarcus’s credit, he was in the room and mentioned his long-standing advocacy for neuro-symbolic approach explicitly.