Garry Tan fine-tuned his own Qwen3.5-397B model in a couple of hours on a Saturday afternoon using Thinking Machines. His framing: fast, usable multimodal is going to enable very mind-blowing personal AI.

The signal is not the benchmark score. It is that a YC president can now personally fine-tune a near-frontier model over a weekend, which means the bar for what counts as a moat in AI infrastructure just moved again.

Garry Tan's GBrain evals show his personal AI stack beating Vector RAG by 38% on LongMemEval and edging out MemPalace by 1%, while being SOTA for reranking and embedding cost, speed, and retrieval success.

The deeper signal: the personal AI layer is becoming a competitive product category with its own benchmarks, not just a hobby project.

Andrew Feldman tells the Cerebras story from a clear first principle: to be radically faster than GPUs—15-20x—you cannot use a similar architecture. Cerebras chose wafer-scale chips the size of dinner plates while everyone else built postage-stamp-sized chips.

For years, nobody cared about speed because AI was a novelty. The turning point was late 2024 into 2025, when models got smart enough for daily use. Speed went from irrelevant to the most important variable, and Cerebras was crushed with demand.

The business arc is dramatic: a $20B+ deal with OpenAI, a deployment agreement with AWS, and a $63B market cap at IPO. Feldman's thesis: just like Netflix went from DVD delivery to movie studio when bandwidth arrived, fast AI will enable entirely new business models that nobody can currently see.

Guillermo Rauch processed 1400 replies to his viral 'show me what you built with AI' prompt. The headline findings: OpenAI is catching up to Anthropic, 'Codex' got more mentions than 'Claude Code', but by model mentions, Claude is still dominating.

The real value of this thread is not the ranking—it is the raw survey data of what actual developers are shipping, which tools they choose, and what products they are proudest of.

The prompt that generated the 1400 replies was simple and powerful: 'Show me the thing you've built with AI you're most proud of. Reply with a working product URL and what model / agent you primarily used.'

With 2195 likes, this is one of the most engaged developer survey threads in recent memory, essentially an organic census of the AI builder ecosystem.

Peter Steinberger built a GitHub dashboard he always wanted: see your repos, open issues/PRs, last released version, commits since last release. And he built it for everyone.

With 676 likes in a few hours, this is a signal that developer tools built with AI agents are crossing from impressive demos into genuinely useful daily utilities.

Amjad Masad shares a user quote that works as a product claim: 'I built my first three apps using Cursor and thought that was fast, but trying Replit for Dial completely blew me away. Built the MVP over a single weekend, and it actually got approved by Apple on the very first try.'

The subtext is clear: the coding tool competition is now measured in speed-to-working-product, and Replit is positioning itself against Cursor head-to-head.

Aaron Levie makes a point that keeps needing to be remade: we constantly confuse task completion with AI with being able to eliminate the whole job. Even as tasks get automated, the definition of the job expands to do more of those tasks, at higher quality, or move on to tasks that haven't been automated yet.

His concrete examples are sharp: the small business that couldn't afford a full marketing agency can now hire a marketer armed with agents who can do as much as an agency did before. The non-tech company that wants to take on larger software projects finally can, and they'll hire to do so.

The one-liner: 'Don't fall into the trap of confusing tasks with jobs.'

Nikunj Kothari is seeing B2B companies finally waking up to an opportunity he wrote about almost a year ago. His advice: don't wait another year. Narratives and vibes are extremely important to stand out amongst the slop.

The implicit point: in B2B AI, the window for establishing your narrative before the market floods is real and finite.

Peter Steinberger shares a practical pattern for working with Codex on big refactors: tell it to maintain a scratch-log with decisions it had to make, tradeoffs, and review fixes, so you can later read through which tradeoffs the agent made and what you forgot to specify.

This is a small but real process innovation—making the agent's reasoning auditable after the fact, rather than hoping it got everything right in a black box.

Aditya Agarwal draws a sharp contrast: six months ago in November 2025, we would mostly just chat with LLMs and be happy about AI. Now in May 2026, these LLMs have produced more code than humans have written over all time.

The framing puts a timestamp on the shift from AI as conversation partner to AI as production engine.

TRQ shares a surprisingly practical use case: running a 'please save me money' prompt on a legacy startup codebase with a community still using it, and it actually works—finding cost saving measures they never had time to look into.

The signal: AI agents are now useful not just for building new things, but for the unglamorous work of maintaining and optimizing old infrastructure that humans never get around to.