X / Twitter
Claude, Anthropic
Code with Claude returns next week. The developer conference has sessions for both developers just starting with Claude Code and teams that have been building deeply with it for some time.
https://x.com/claudeai/status/2050252933866930339
Aaron Levie, Box CEO
Levie systematically outlined how enterprise-software business models will evolve once agents become software's largest users. First, human seats will remain, but each must include an allowance for agents to use APIs on the user's behalf or become obsolete. Second, agents may have their own seats for stateful, long-term work, but pricing must differ entirely: one customer might use 1 agent for everything, another 1,000, making per-head pricing largely unworkable. Third, headless usage beyond seat allowances will mostly use consumption pricing similar to PaaS, though future APIs may charge per outcome rather than a sequence of calls. Levie also considered Atlassian's better-than-expected earnings unsurprising: with 100 times as many agents as humans working in systems, volumes of code, contracts, invoices, and designs will surge. The underlying systems managing this work, including security, compliance, and workflows, will only grow more important.
https://x.com/levie/status/2050051426446152159
https://x.com/levie/status/2050295657836277764
https://x.com/levie/status/2050240083325030404
Garry Tan, Y Combinator CEO
Garry updated his comparison of GBrain, his open-source personal knowledge-graph tool, with Mempalace. They looked similar three weeks ago but have now clearly diverged: GBrain better suits OpenClaw/Hermes personal AI rather than precise needle-in-a-haystack retrieval. He recommends installing GBrain alongside a Karpathy-style knowledge wiki for substantial compounding benefits when using Hermes Agent or OpenClaw.
https://x.com/garrytan/status/2050095919157350644
https://x.com/garrytan/status/2050096324100682097
Builder Zara Zhang
Zara shared a shift in collaborating with coding agents. Many people treat them as employees and issue orders; she treats hers as a co-founder, presenting a problem, describing the situation, and asking for its opinion. She believes this produces better results.
https://x.com/zarazhangrui/status/2050326543797469415
Dan Shipper, Every CEO
Dan noted an interesting contrast: models know more than any individual human, yet any individual human learns faster than models. This tension shapes the deeper logic of human–AI collaboration.
https://x.com/danshipper/status/2050304359024759242
Nikunj Kothari, FPV Ventures Partner
Nikunj has just built and launched a paid project with Railway, Conductor, and Claude, his first revenue-generating side project. Each report costs $8–9, with sales essentially at cost and a half-price discount code offered to early users in exchange for feedback.
https://x.com/nikunj/status/2050353986742698400
Podcasts
No Priors — Baseten CEO Tuhin Srivastava on the AI Inference Crunch, Custom Models, and 30x Scale
Key takeaway: Baseten CEO Tuhin Srivastava says the company grew 30-fold over the past year and expects revenue above $1 billion this year. Customer-customized models account for 95% of inference tokens. Large-scale enterprise adoption remains early, and combining post-training with inference infrastructure will be crucial.
Tuhin Srivastava founded and leads Baseten, an AI inference cloud platform. It provides inference infrastructure for AI-native companies, including leading applications Abridge, Cursor, Open Evidence, Decagon, and Clay, focusing on deploying and serving high-performance custom models. Baseten grew 30-fold over the past year, with 2026 revenue expected to exceed $1 billion.
Why the Inference Market Is Exploding
Tuhin sees the biggest change of the last 24 months as open models crossing capability thresholds while post-training techniques such as RLHF and RLAIF mature enough for enterprises to realize they can own their inference capabilities instead of relying on one closed API. This creates a long tail of deployed custom models, directly benefiting infrastructure providers such as Baseten.
Is There Room for an Independent Application Layer?
The host asked whether labs would absorb independent applications. Tuhin replied that real defenses lie not in model choice but in unique user signals companies can obtain, accumulate, and encode into workflows. He cited Abridge, a clinical-note AI assistant deeply integrated with most US hospitals. Its unique clinical-workflow signals are inaccessible to frontier-model companies. At sufficient scale, these signals support post-training proprietary models and create longer-term advantages. By inference-call volume, he believes 99% of the market has yet to actually deploy AI, leaving enterprise adoption extremely early.
95% of Tokens Come from Custom Models
A striking figure: more than 95% of Baseten's served tokens come from customer-customized models, not unchanged public weights. Customization includes domain-specific fine-tuning to improve quality and quantization and compilation optimizations to improve inference performance. Tuhin said customers always begin with "can the strongest model solve this?" before optimizing cost. This explains their diverse models, from GPT OSS, DeepSeek, and Moonshot to the TTS model Orpheus, following a "frontier first" strategy.
Acquiring Parzed to Expand into Post-Training
Baseten acquired the post-training team Parzed a few months ago. Parzed had been a Baseten customer, training and running custom models on its infrastructure. Tuhin explained the logic: more infrastructure customers need post-training support, and post-training and inference are two sides of one problem. Model quantization, how training affects quantization strategy, and the complete loop of "inference generating data, evals identifying reward signals, and further post-training on those signals" naturally connect the businesses. He advises startups to prove product value with the best closed model before customization: "post-training before PMF" is the same mistake as "buying GPUs before PMF."
Geopolitics and Chinese Open Models
Asked about Chinese open models such as DeepSeek, Tuhin said he finds it hard to see how network-isolated models threaten data security and has seen no evidence of actual technical backdoors. He considers a strong US open-model ecosystem strategically essential, but if the United States cannot provide one, temporarily building with DeepSeek as though it were "a model from Meta" is pragmatic. He highlighted a geopolitical paradox: by subsidizing open models, China's government indirectly lowers US companies' AI costs and, in a sense, subsidizes American innovation. The host added that the strongest frontier capabilities still belong to closed models from Anthropic, OpenAI, and Google.