Y Combinator CEO Garry Tan
Y Combinator CEO Garry Tan quotes a result claiming that 3 million tokens of advice from ordinary people facing moral dilemmas reduced malicious behavior to 0%, with a 28-fold efficiency gain across model generations; the batch does not detail evaluation conditions. He separately advocates eval-driven development loops for agent skills, removing components that newer models no longer need. He says he removed /investigate because Opus 5.5 did better without it. He also praises Aside’s computer use and sees a genuine Chromium browser as less likely to be blocked by websites. These are a quoted finding, personal experience and product assessment, not proof of permanently harmless models or permission to bypass access restrictions.
AI Engineer organizer Swyx
AI Engineer organizer Swyx plans weekly coverage of Anthropic’s potential IPO until it does or does not happen. He cites experience covering major IPOs including BABA and TEAM and plans to combine an equity hedge-fund analyst perspective with AI engineering, alongside Jamin Ball. This announces coverage rather than establishing that an IPO has occurred or is certain.
FPV Ventures partner Nikunj Kothari
FPV Ventures partner Nikunj Kothari explicitly labels a fundraising narrative as satire: raise $2M, spend $200k on a launch video, $200k promoting it and $500k on tokens, then package friendly customers’ LOIs and cARR as revenue. He describes tranches moving a valuation from $40M to $80M and $100M, late incorporation to dramatize rapid revenue growth and $100k spent shaping a vision. He says the individual scenarios have occurred in his experience; this is not an actual startup plan or operational recommendation. Another post questions startups presenting paid partnership exposure as organic X views. He also reintroduces texts wrapped, which analyzes iMessage, WhatsApp and phone calls and claims local, no-AI, open-source processing on Mac with refreshed design, insights and accuracy. Access is through a terminal command or Claude Code/Codex and requires full disk access; the local-processing claim does not remove the breadth of that permission.
Builder Zara Zhang
Builder Zara Zhang argues that a personal website is the new resume. LinkedIn primarily lists titles, whereas a personal site can demonstrate taste, thinking and actual work. She encourages people not to be confined by existing social-media formats.
AI tutorial author Peter Yang
AI tutorial author Peter Yang observes that apps supporting models from multiple labs, including Conductor and Cursor, seem to be getting acquired by the labs themselves. He points to remaining apps T3 Codes and Orca as open source. He believes users and enterprises still want multi-lab applications; both the acquisition pattern and demand outlook remain his observations and judgments.
SPC general partner Aditya Agarwal
SPC general partner Aditya Agarwal uses Starship’s development to argue that ambitious technology requires both urgency and patience. The post does not detail development stages or a schedule.
Former Claude Code contributor Thariq
Thariq, formerly of Claude Code, highlights creator apps such as Instagram Edits and Roblox Studio shipping built-in MCPs so agents can do some of the work. Power users can still collaborate and continue in familiar tools. He values continuity between agent assistance and existing professional interfaces rather than a separate working environment.
Vercel CEO Guillermo Rauch
Vercel CEO Guillermo Rauch describes generative UI for conversational agents: combining AI SDK, Chat SDK and json-render to move from LLM output to interfaces on platforms such as Slack and Teams. He expects this to become a dominant Human-Agent Interaction, or HAI, pattern. The post provides no cross-platform coverage or adoption figures supporting that forecast.
Latent Space — Controlling Biology - Fabian Theis, Helmholtz Munich
Key takeaway: Progress in virtual cells requires well-chosen prediction tasks, strong simple baselines and experimental feedback as well as larger models.
Fabian Theis of Helmholtz Munich studies single-cell modeling and contributes to research tools including Scanpy, AnnData and scverse. He cautions that perturbations include not only gene knockouts but drugs, culture conditions and cellular signals, and virtual cells should not be equated with perturbation prediction alone.
The task must justify model complexity. He says logistic regression can remain strong for cell-type classification within an organ, whereas cross-organ classification or predicting neighboring cells’ expression from one cell’s state is harder. His instruction to students is: “I know you wanna do the most fancy deep learning thing, but please, please, you know, have the baseline.” Noisy drug data can make even predicting the mean difficult to beat, and holdout design and metric selection can alter comparisons.
Representing state differs from predicting intervention. Learning a cellular state space with a foundation model does not establish knowledge of an action’s consequences. The latter resembles a world model taking state X and action A to predict the next state X′. Theis describes combining representation learning with generative methods such as flow matching and calls for demonstrated scaling behavior rather than simply copying NLP training objectives.
The next step is guiding real experiments. He introduces Lab Compass, an inverse generative model intended to work backward from a desired cellular state to experimental inputs. He reports three or four lab-in-the-loop iterations with Cambridge collaborators. This requires experimentalists to invest time testing model suggestions, not merely a loop in a paper’s diagram. Current work is closer to active learning; reinforcement learning across the whole process remains a vision, with experimental cycle times constraining automation.
Organization matters too. The Munich School for Data Science’s Y-shaped supervision brings a methods PI and a domain PI together on projects. Students attend both lab meetings and turn incompletely formulated biological questions into measurable tasks. Commercially, he sees standalone biology foundation-model platforms often shifting toward drug assets and finds combinations of modeling and distinctive data generation more promising. Research results, industry observations and forecasts remain attributed to him. Organoids retain limitations such as missing immune systems; these claims do not establish complete organ simulation or clinical efficacy.