Unique Research · F2AI Focus · March 13, 2024
Adaptive — the startup spun from the Falcon AI model team — announced on March 11, 2024 a $20M venture round to help enterprises customize LLMs to their own needs.
Co-founders Julien Launay, Daniel Hesslow (chief research scientist), and Baptiste Pannier previously built Abu Dhabi’s Technology Innovation Institute (TII) Falcon open LLM series; several engineers also from Hugging Face and AWS. Seed led by Index Ventures; ICONIQ Capital, Motier Ventures, Databricks Ventures, HuggingFund by Factorial, and angels. Valuation undisclosed; The Information had cited ~$100M.
Productizing GenAI alignment for enterprises
Consumer internet (TikTok, Netflix, X) felt “personal” because of recommenders. GenAI can scale machine content, but replies often lack human warmth. OpenAI and DeepMind use preference tuning / alignment so models feel more personalized — yet most enterprises lack the expertise and infra, so they ship constrained GenAI experiences.
Adaptive productizes alignment so enterprises can deploy models that are not only correct but feel personal. Focus: improving RLHF (the step that turns next-token predictors into chatbot-grade systems). Classic RLHF hires agencies to rate responses — costly, and feedback quality often misses business needs. CEO Launay: getting models to behave as intended is hard.
Adaptive’s bet: let LLMs continuously learn from employees’ and customers’ software interactions for richer training signal; capture interaction patterns; train/fine-tune accordingly. A platform simplifies the RL stack so non-expert teams can pick data, objectives, and algorithms. Also exploring RLAIF (one AI rates another) to cut cost and widen training data.
Works with any open-source LLM or in-house model — not third-party proprietary fine-tunes. Platform also helps A/B LLMs, monitor post-deploy behavior, and map LLM outputs to business KPIs. Customers already onboard; new capital expands GTM/sales.
Falcon LLM — open-source firepower from Abu Dhabi
Falcon is released by UAE’s TII under Abu Dhabi’s Advanced Technology Research Council (ATRC). Sep 2023: Falcon 180B trained on up to 4,096 GPUs via Amazon SageMaker on 3.5T tokens (~7M GPU-hours) — ~2.5× Llama 2 size, ~4× compute. Open LLM Leaderboard (Nov 2023 refresh) put Falcon roughly on par with Llama 2 70B (~67.85 under updated benches). Open weights invite inspect/modify/redistribute; NVIDIA + Microsoft Azure cited as infra partners for enterprise deploy.
Middle East political and capital ambition to shape global foundation models is explicit — ATRC/TII frame open collaboration as the platform. Open models democratize access (researchers to enterprises without proprietary license fees) but raise IP, quality-control, and misuse risks with many contributors.
Industry uses cited: machine translation, text generation, semantic search, sentiment analysis.
Is open source the main path?
Open models pool collective improvement — and complexity/security risk. Closed models offer tighter QC and vendor support but limit customization and create vendor lock-in. Open wins on data privacy (run on private servers) and customization; choice still depends on enterprise needs.
At Unique Research’s AIGC forum, OpenCSG founder Chen Ran argued foundation models should be open to customers — customer data IP implies customer-owned trained models; open source is commercial-model and industry-chain critical. Dell AI SVP Matt Baker: large public models alone have “almost no value” for private enterprise — too generic/bloated, hard to access own data. Zhipu CEO Zhang Peng: open and closed are complementary, not opposites — open keeps community diversity; closed commercial builds serve security, cost, efficiency, sustained value.
2024’s open-source AI race is only beginning.
3 CN images; attach on ingest.
— END —