Harness-1 Open-Source Search Agent Competes with Top Models; Over 25 Major Open-Source Releases This Week; AI Unicorn Landscape Emerges

Key Takeaways

1. Harness-1, a 20B open-source search agent, achieves frontier performance in long-horizon search, competing with Opus-4.6 and surpassing GPT-5.4, with costs and latency similar to Context-1.
2. Google's Gemini Pro iteration lags behind Claude and GPT, widening the performance gap despite Flash 3.5 improvements.
3. This week saw over 25 major open-source weight releases across LLMs, image generation, audio, vision, and 3D/video.
4. 21 AI startups now exceed $10B valuation and $100M revenue, including Crusoe, ElevenLabs, Mistral, Perplexity, DeepSeek, xAI, OpenAI, Anthropic, and more.
5. California's non-compete law enables knowledge spillover that surpasses GitHub, arXiv, and Hugging Face combined; researchers leaving with tacit knowledge have raised over $100M.
6. Intel partners with Aravind Srinivas to bring local model and hybrid inference PCs to Ultra Series 3 laptops.

1. AI Model Releases and Performance

1. Hugging Face highlighted Harness-1, a 20B search agent trained with state externalization, achieving frontier-level long-horizon search. It rivals Opus-4.6, beats GPT-5.4, and keeps costs and latency at Context-1 levels. The model is open-source. — via 1
2. Ethan Mollick noted that Gemini Pro iteration speed is far behind Claude and GPT, with the last update being 3.1 Pro in February. This performance gap is widening, and the Gemini 3.5 Flash model cannot compensate. — via 1
3. Yann LeCun summarized this week as the most insane week for open-source AI model releases, listing over 25 major open-weight releases covering LLMs, image generation, audio/speech, vision/VLM, and video/3D world models. — via 1

2. AI Industry Landscape and Policy

1. Deedy shared a list of 21 AI startups with valuations over $10B and revenue exceeding $100M. Notable names include Crusoe, ElevenLabs, Mistral, Perplexity, DeepSeek, xAI, OpenAI, and Anthropic. Some valuations are reported as unverified. — via 1
2. Swyx argued that California's non-compete law impacts knowledge dissemination more than GitHub, arXiv, and Hugging Face combined. Researchers can leave with protected tacit knowledge, raising over $100M in funding, leading to a decline in research papers and lab publications. This inspired him to found @aidotengineer as a product-focused industry conference. — via 1

3. AI Hardware and Research Direction

1. Aravind Srinivas delivered a keynote at Computex 2026 outlining Intel's vision from silicon to system to software, covering PCs, edge, data centers, and intelligent centers. He also announced a partnership with Intel and @LipBuTan1 to bring local model and hybrid inference PCs to Intel Ultra Series 3 laptops. — via 1 2
2. Yann LeCun advised computer vision researchers transitioning to robotics to focus on real sensorimotor problems, especially hand-object interaction, contact and force, proprioception, and tactile sensing, and to be wary of cherry-picked demos. — via 1