Alexander Bergman

Unpopular opinion: The RL hype is a red flag for technical depth. When I hear founders or VCs casually dropping "reinforcement learning" in pitches, it's usually code for "we don't actually understand our problem space." Here's the reality check: RL only works when you have: - Perfect simulation environments - Extremely fast iteration cycles - Clear reward functions - Tolerance for massive compute costs That's why it succeeds in games (perfect simulation) and search optimization (Massive compute capacity at Deepmind AlphaGo/Anthropic/OpenAI) but fails spectacularly in the messy, slow-feedback real world. The graveyard of "RL-powered" startups tells the story. They burned through $$$$ trying to apply algorithms to problems that needed simple heuristics or supervised learning. Most of the time, a well-tuned evolutionary algorithm will outperform your fancy RL setup while using <1/10th the compute and delivering results you can actually explain to stakeholders. The next time someone pitches you an RL solution, ask them: "Have you tried solving this with a basic optimization approach first?" Watch how quickly the conversation changes. I am surprised when people are surprised that Direct Preference Optimization (DPO) can exceed PPO-based RLHF [1] or Reflective Prompt Evolution Can Outperform Reinforcement Learning [2]. [1] https://lnkd.in/gCY3bcTQ [2] https://lnkd.in/gQRwyQyM

45

2 Comments

Axiom, a pre-product AI startup founded by a Stanford math PhD Carina Hong, is raising at a $300m valuation. Goal is to build AI trained on math proofs that can solve complex math problems with certainty, for Wall St. quants. Carina graduated from MIT with a double major in mathematics and physics. She then earned a neuroscience master’s degree with dissertation distinctions from Oxford, now pursuing a law degree and PhD in mathematics at Stanford.

63

3 Comments

I'm extremely excited to announce that SAI is now open to the public! We’ve spent the past two years building this in stealth, designing a platform where reinforcement learning researchers/engineers/enthusiasts can submit models, compete in structured challenges, and learn from each other’s work. Today marks a big step forward because we're inviting the entire RL community to help shape what comes next. More competitions and features are on the way. But for now: - Explore what’s live at https://competesai.com/  - Give us your feedback - Share it with someone who might find it valuable We’re building this in public, and with your input, we hope to make SAI a home for reproducible, competitive RL. Documentation: https://lnkd.in/euczcE2p #AI #ReinforcementLearning #RL #MachineLearning #Robotics

81

6 Comments

⚠️ Founders - consider the below reasons to not use Forward Deployed Engineers (FDEs) to acquire your *first* customers. This signals (1) lack of PMF; (2) miscalculation b/w development and sales timeframes; and (3) mischaracterization of Palantir Technologies timing when they deployed their FDE model. 💡 At the early-stage, it's crucial to demonstrate ability to repeatedly acquire customers with limited resources - capital, team, & time - that supports the narrative of 'more funding will amplify my traction.' Bespoke solutions for one company demonstrates zero repeatability. These custom solutions will not carry over to other customers within the same cohort. 💡 AI accelerates engineering but sales, system integration, and stakeholder buy-ins are perpetually slow. This imbalance results in concentrating all your engineering resources into one company, ignoring other opportunities, and still not going live on an accelerated timeline. 💡 Palantir didn't start using FDEs until year 3 when they got their first *gov't* customer, not their first customer. Gotham was already built and had strong PMF. Gotham allowed Palantir to take GTM risk by using FDEs b/c if it didn't work, Gotham was humming in the background. TLDR: At the early-stage, repeatable customer acquisition is the holy grail. Don't let AI deceive you that it can accelerate human behavior like it has with development cycles. It C.A.N.N.O.T.

32

1 Comment

Nano Banana Hackathon hosted by Google DeepMind and Cerebral Valley in SF, with fal and ElevenLabs. The bar was incredibly high. Believe it or not, the 1st prize winners (Banana Peel) built a social network for editing photos through threaded comments in just a few hours, with Nano Banana generating images live. • What’s Nano Banana? Google’s advanced image generation and editing model (officially Gemini 2.5 Flash Image), delivering top-tier quality at a low cost: just $0.039 per image. • High-Speed Outputs: Not only affordable, it’s also blazing fast—generating 1024x1024 images in ~2.3 seconds on cloud, using 2.1GB GPU memory with 15% better energy efficiency than competitors. Expect it to run on TPU-powered Android phones soon. • Multimodal: Built on Multimodal Diffusion Transformer (MMDiT), it processes text prompts and multiple images together in a single workflow, similar to gpt-image-1 (chatGPT’s image generation). This enables seamless image merges (e.g., blending a person, clothes, and scene), style transfers, and consistent character details with accurate lighting and perspective. Try it out for free on Google AI Studio, or as a third-party model on fal. Ivan Porollo Paige Bailey Burkay Gur

41

2 Comments