Rorry Brenner

Our latest ImageNet experiment found perforating a ResNet matched the accuracy gain of collecting 25% more data at the 1M-image mark. Excited to see this hold on the foundational computer vision benchmark.

More data = better accuracy is a core tenet of modern ML training. But accuracy scales logarithmically with dataset size (Sun et al., 2017), so every additional datapoint buys less than the last.

Full writeup: linkedin.com/pulse/best-mod…

Why it works: distillation changes what the student learns from. Perforation changes how it learns. Different axes, compounding gains.

Model optimization techniques aren't competitors. They stack. We combined knowledge distillation with dendritic learning and nearly doubled the accuracy gain. 🧵

🇺🇸 The biggest AI conference of the year takes place next week in San Francisco: AI Engineer World's Fair. Presenting companies include OpenAI, Anthropic, DeepMind, and many other leading AI companies. For those visiting, we put together a short list of the best events happening around the conference week. (1) Jun 29th | Reading Group: Agents' Last Exam | By @SnorkelAI luma.com/4qyhs0cw?utm_s… Berkeley researchers present Agents' Last Exam, a benchmark for testing how AI agents handle real work tasks across different fields. (2) Jun 29th | Artificial Analysis Intelligence Index | By @ArtificialAnlys @AICollectiveCo @WandaWang007 @grmcameron @_micah_h luma.com/qdl9mr2e?utm_s… ​Join Artificial Analysis, Google, NVIDIA, and Z ai for a conversation on how teams compare AI models beyond benchmark scores, including quality, speed, and cost. (3) Jun 29th | The Rise of Support Engineering & AI | By @plainsupport luma.com/jgggw759?utm_s… Support engineering leaders from Mintlify, Bland, and Plain discuss how the role has shifted as agents take on first-line work, what gets harder, and how teams are restructuring. (4) Jun 30th | Agents in Production: The Full Stack | By @SurrealDB @tobiemh @MarcBaiza luma.com/hqz8wi79?utm_s… SurrealDB, AWS, and Vercel engineers cover persistent agent memory and production data infrastructure. (5) Jun 30th | ML Suds & Science | By @pebble_bed @nicilevv @rorrybrenner luma.com/86hownrw?utm_s… A meetup for people in the trenches of ML training runs, optimizer choices, and architecture decisions that get lost between arxiv abstracts and production. (6) Jul 1st | What open agents can actually do: a happy hour with Together AI & NVIDIA | By @togethercompute luma.com/AIEWF-OpenAgen… Together AI and Nvidia hold a fireside chat on collective agent intelligence, the EinsteinArena and DSGym work, and what open models like Nemotron need to power agentic systems. (7) Jul 1st | Corgi Cafe Arcade | By @QodoAI @DanaFine7 luma.com/n181x4b1?utm_s… Qodo hosts a full-day pop-up at Corgi Cafe with coffee, food, arcade games, and a live demo, plus a Nimble-sponsored breakfast and afternoon games with prizes. (8) Jul 2nd | AI Engineer World's Fair Rooftop Afterparty w/ Apify × Massive × TopK | By @apify @jerguslejko @jsongrad @jancurn luma.com/aie-sf-afterpa… ​A rooftop evening after AI Engineer World’s Fair with builders and operators working across web data, search, and infrastructure for AI apps and agents. Dozens of more events here luma.com/genai-sf

Bringing together my loves for startups, comedy & community 🎤✨ This Monday, is the first-ever Pitch Nite at Devil's Canyon: part pitch competition, part stand-up, part Speed Friending. 120+ RSVPs - come join us! 🗓️ SF 🔗 lu.ma/peqqlfpc #PitchNight #Comedy #startups