Thariq
Thariq revisited his attempts to make videos with Claude Code a year earlier. Those pieces were not finished in one generation. He had to point out incorrect details repeatedly and spend substantial time iterating. Looking back now, he sees a large improvement in capability. His reflection places that progress in the cost of revisions during creative work.
Garry Tan
Garry Tan shared his current favorite way to fix bugs. He combines capy with GStack’s /autoplan. The demonstration addresses a production issue. His model choice is GPT-6 at medium reasoning. He presented the combination as an actual repair workflow rather than a standalone code-generation demo.
Peter Yang
Peter Yang is building a conversational Japanese app with the Gemini audio API. It contains 10 lessons with 10 phrases each. To try Google’s new audio APIs, he used Antigravity for the first time and asked where to send feedback. That day, he also said his experience of Claude limits had shifted from barely usable to very generous. These observations came from hands-on tool use while building a learning app.
Dan Shipper
Dan Shipper first wrote a novelization of Plato’s Protagoras. He then asked Opus 5.5 to turn that adaptation into a film. His shared work includes scenes one and two. The two posts show consecutive pieces of the same creative project. The workflow connects literary adaptation with film generation.
Guillermo Rauch
Guillermo Rauch called for rejecting AI output produced without understanding. He worries that a constant stream of low-quality, unverified prose will exhaust readers and discourage careful reading. He cited a discussion that described algorithm and data-structure improvements as compiler improvements, showing how the problem reaches technical explanations. Low-quality output is not limited to code and PRs. Rauch wants AI to improve understanding and creativity rather than make readers bear the cost of sorting out incorrect accounts.
No Priors | Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon
Core takeaway: Diffusion language models begin with parallel inference, delivering answers of comparable quality faster and more efficiently.
Inception co-founder and CEO Stefano Ermon has researched generative models at Stanford since 2014. In 2019, he and Yang Song explored generation through iterative denoising. A key result arrived in 2024: at GPT-2 scale, diffusion matched autoregressive perplexity on the same data and parameter count while generating text about 10 times faster. He then founded Inception to commercialize the approach.
First, sequential generation is a bottleneck. An autoregressive model must generate earlier tokens before later ones, spending considerable GPU time moving weights. Diffusion processes multiple tokens together and repeatedly refines a coarse output, mapping more closely to parallel matrix operations. Ermon’s view is that “the more parallel solution is the one that is eventually going to win.”
Second, speed already has clear uses. Voice agents combine speech recognition, LLM reasoning and tool calls, and speech synthesis; delays anywhere affect the experience. Ermon described Open Call moving from models on specialized chips to Mercury, obtaining the required speed on NVIDIA GPUs. Mercury retains an OpenAI-compatible API and structured outputs so customers can reuse their existing stack.
Third, research needs a complete serving system. Inception has roughly 50 people and builds its inference engine, post-training tools, and evaluations as well as models, iterating with real customer feedback. Existing autoregressive serving frameworks cannot directly meet its production requirements. That engineering work becomes part of the company’s accumulated capability.
Fourth, speed does not mean frontier intelligence has already been reached. Ermon explicitly noted that many workloads still need stronger models; he estimated that 20% to 30% of workloads are particularly latency-sensitive. Better controllability and data efficiency remain research directions. His initial focus is measurable speed that customers are willing to pay for. He also acknowledged an immature tooling ecosystem: deployment and on-premises use are harder, requiring the team to build much of the supporting infrastructure itself.