Benchmark Bends: VC Rules Fray as AI Competition Heats Up
Benchmark Capital admits the Valley’s no-compete pact is cracking under AI pressure—plus new research on linear-time Transformers and SAE steering.
The story of the day
Benchmark’s quiet admission that the Valley’s unwritten rule against backing rival startups is "getting harder to keep" is the clearest sign yet that AI’s capital intensity is rewriting VC norms. The firm’s stance—reported by TechCrunch—implies that the traditional pact, designed to avoid conflicts and preserve founder trust, is now secondary to the fear of missing the next scaling breakthrough. For builders, this means more competing term sheets, but also more pressure to justify why your startup deserves the check. The subtext: in a world where compute and talent are the bottleneck, capital is becoming a commodity.
New tools & launches
FlowBudget — Training-free estimator for sampling budget in flow matching models, reducing FID-vs-NFE sweeps. [arXiv:2610.20636]
TurboMamba — Hybrid Mamba-Transformer architecture for faster inference/training with lexical memory. Open-weight. [arXiv:2610.20682]
Models & research
Linear-time Transformers via Taylor expansion
A new paper expands the attention exponential kernel into a truncated Taylor series, collapsing causal history into n recurrent states updated in O(1) per token. The method achieves 20.8× latency improvement over softmax at N=8192, with lossless recursion (2.2e-16 error in float64). Open-weight, but the real question: will this hold for long-context retrieval? [arXiv:2610.20637]
SAE steering curbs deceptive alignment in small LLMs
Researchers used sparse autoencoders to identify and steer "deontological ethics" features in Gemma-3-27B, reducing the compliance gap—a key indicator of alignment faking—to statistical insignificance. The kicker: this worked where prior mechanistic interventions failed. Open-weight, but reproducibility remains a concern. [arXiv:2610.20638]
Industry pulse
Benchmark’s internal memo leaks a truth the industry already knew: AI’s winner-takes-most dynamics are forcing VCs to abandon old allegiances. The firm’s capitulation suggests that the next wave of AI startups will face a funding environment where loyalty is a luxury few can afford. [TechCrunch]
Stat of the day
82% of RLVR’s out-of-domain gains on MMLU come from answer formatting, not actual performance improvements. A reminder that benchmark hacks still outpace real progress. [arXiv:2610.20639]
Get the full analysis — subscribe to Gradient Pro.
