Thursday 4/16 ‣ Agent Builders Breakfast - Founders & Builders in SOMA, SF luma.com/xmv1m6qt @miradu ‣ AI Breakfast: Build your AI workspace in one morning luma.com/aibreakfast2 @rajoshighosh ‣ Platform Engineering & AI luma.com/intuitossmtvapr2026 @cbergman ‣ Voice AI Meetup: Medical Mode luma.com/609tv1po @ryanseams @theaievangelist ‣ Claws Out🦞 GMI ClawHub Demo Night luma.com/tklag2kv @nicoleegong @gmi_cloud @yuqih ‣ Voice AI builders night luma.com/voice_builders @modal @braintrust ‣ AI Meets HumanX Social w/ Rootly AI, MongoDB, Runpod, & More! luma.com/n22qf90w ‣ Slides Down: AI Founders Party by DigitalOcean & Zendesk Apr 16 luma.com/jzopt9nv @neffko @julianachyzhova @yfilipch

Releasing a new "Agentic Reviewer" for research papers. I started coding this as a weekend project, and @jyx_su made it much better. I was inspired by a student who had a paper rejected 6 times over 3 years. Their feedback loop -- waiting ~6 months for feedback each time -- was painfully slow. We wanted to see if an agentic workflow can help researchers iterate faster. When we trained the system on ICLR 2025 reviews and measured Spearman correlation (higher is better) on the test set: - Correlation between two human reviewers: 0.41 - Correlation between AI and a human reviewer: 0.42 This suggests agentic reviewing is approaching human-level performance. The agent grounds its feedback by searching arXiv, so it works best in fields like AI where research is freely published there. It’s an experimental tool, but I hope it helps you with your research. Check it out here: paperreview.ai

aidan bench update: i ran llama 3.1 405b at bf16 (shoutout to @hyperbolic_labs) and we got a *way* better score. 405b fp8 is around gpt-4o-mini-level 405b bf16 beats claude-3.5-sonnet give me bf16 or give me death

@felipehoffa @cbergman so great to see you at the @awscloud GenAI Loft today!

Self-care life hack: if you feel a bit down/tired, paste the url of your website/linkedin/bio in Google's NotebookLM to get 8 min of realistically sounding deep congratulations for your life and achievements from a duo of podcast experts 😂

My take on "LoRA Learns Less and Forgets Less" 1) "MLP/All" did not include gate_proj. QKVO, up & down trained but not gate (pg 3 footnote) 2) Why does LoRA perform well on math and not code? lm_head & embed_tokens wasn't trained, so domain shifts not modelled. Also reason why "LoRA Forgets Less". Use "modules_to_save" in HF PEFT or "lm_head", "embed_tokens" in @UnslothAI 3) Code rank=256 used α=32 (too small!) (pg 18), but Maths α=2*r=512. RS LoRA paper showed α/sqrt(r) needed for larger ranks. & common practice is 2*r. So also why Code did worse than Maths 4) Extrapolating Maths vs fft looks good. Small datasets LoRA>fft, but I theorize that's because of reason 2 5) LoftQ & PiSSA paper init LoRA from SVD(W) => papers show comparable perf of LoRA 6) LoRA paper shows B matrix needs larger lr. DoRA (mentioned in paper) learns these scalars. TLDR: Code worse since α=32 is too small. No embed_tokens, lm_head (or layernorms), not even gate_proj? Better init & lr scaling can help For continued pretraining, I advise people to train on all layers (inc gate) lm_head, embed_tokens, use RS LoRA, use rank>=256 LoRA paper: arxiv.org/abs/2405.09673 RS LoRA paper: arxiv.org/pdf/2312.03732 LoRA paper: arxiv.org/pdf/2402.12354 PiSSA paper: arxiv.org/pdf/2404.02948 DoRA paper: arxiv.org/pdf/2402.09353