What makes early-stage AI accelerators useful (and what doesn't)?

Hi HN — we recently launched the Berkeley Xcelerator (rdi.berkeley.edu/xcelerator), a non-dilutive accelerator program run by Berkeley RDI (rdi.berkeley.edu) for pre-seed and seed-stage teams building in AI and agentic AI. We’d love to get some feedback from the community!Over the past three years, Berkeley Xcelerator has supported 110+ teams across AI, cybersecurity, and decentralized technologies, whose founders have gone on to raise $650M+ in follow-on funding, spanning 100+ countries. Some concrete details about the Xcelerator itself: • The program is non-dilutive (no equity taken) • Open to pre-seed and seed-stage AI / agentic AI startups • No UC Berkeley affiliation required • Selected teams receive support through Berkeley RDI’s research community and ecosystem partners • Enablement includes cloud, GPU, and API credits from industry partners (including Google Cloud, Google DeepMind, OpenAI, and Nebius, with more to be announced) • The program culminates in a Demo Day at the Agentic AI Summit (Aug 1–2, 2026) at UC Berkeley, where we are expecting 5,000+ in-person attendees Here’s what we’d really like input on: • If you’ve built or joined an early AI startup, what actually helped you most early on? • If you’ve done an accelerator, what helped and what was a waste of time? • For technically deep projects (infra, agentic systems, safety-sensitive work), what kinds of feedback or structure mattered most before product-market fit? If you’d like to apply to the Berkeley Xcelerator, applications are open through the end of February. (forms.gle/KjHiLAHstAvfHdBf7)