Swyx connects Noam Brown's argument to the open vs closed model debate: if inference budget is what matters, then open models — which have much better dollar-per-token mileage — should report performance by dollar cost, not token count.

This is a strategic insight for open model providers: reporting by dollar cost on popular inference providers would make their models look much more competitive against frontier API models, because you get more thinking per dollar.

It also pressures closed model labs to be more transparent about how much test-time compute their reported benchmark numbers actually used.

Levie's thread on token cost optimization is really a thesis for applied AI companies: the layer between the model and the actual work needs to deeply understand the customer's domain, workflows, and business processes.

His argument: enterprises get more intelligence per dollar when an applied AI company has already done the work of evaluating models for specific use cases, tuning the UX, supporting adoption, and building optimal architecture.

This is effectively the business case for vertical AI: horizontal plays can't go deep enough into any single workflow to optimize cost and performance the way a specialist can.

Rauch's security warning is significant: Mythos/Sol cybersecurity capabilities are dual-use — equally effective offensively and defensively.

His concern: if adversaries get an equivalent offensive capability, it's a serious threat to US companies that haven't found their latent vulnerabilities yet.

His recommendation: run deepsec or similar harnesses with available frontier models to proactively find and fix vulnerabilities before adversaries do.

1225 likes for a security recommendation from a CEO who runs a deployment platform suggests the builder community takes this seriously.

Steinberger's quote — 'History teaches us that access blockage rarely stops determined users' — got 670 likes and 64 retweets.

Combined with his Apple Car post (530 likes) and his monitor setup deliberation (341 likes), he's having a particularly engaging day of observations.

The access blockage quote likely refers to API restrictions or feature gating — and historically, determined developers find workarounds, which is both a challenge and an opportunity for platforms.

Sottiaux shipped a substantial Codex update: smooth handling of super long threads, a hoverable navigation rail, expanded settings search, zoom-level fixes across tooltips/dialogs/menus, Slack-paste preserving markdown — and most importantly, a dedicated Pets panel.

1383 likes and 170 replies make this the most-engaged Codex product update post yet — the Pets panel joke clearly resonated.

The substance matters though: the long-thread performance fix and the Slack markdown paste are real pain points being solved. These are the kind of polish updates that turn a tool from 'impressive' to 'daily driver.'

Zara's one-liner — 'Last year I barely knew how GitHub worked. Now I have 10k followers on GitHub (and still can't write code by hand)' — got 2149 likes and 151 replies.

This is the most relatable AI-era builder story: someone who isn't an engineer, but who can connect technology with user problems, use AI tools to build, and tell stories about the products.

Her follow-up is important: 'I'm not an engineer; these are all side projects I built for fun. I just love connecting technology with user problems.' This is the Builder PM archetype from earlier discussions, made real.

2149 likes suggests this resonated enormously — it's the proof that the barrier to building has dropped far enough for non-engineers to create real value.

Rauch bought rauch.com — 1502 likes for a CEO expressing simple joy about owning his own name domain.

His framing: 'Domains are awesome. The internet is awesome.' — a reminder that even in the age of agents and AI, the fundamental primitives of the internet still matter.

The children's clean email handles line is sweet — it's a CEO thinking about digital legacy, not just product strategy.

Matt Turck's smart glasses history is comedy gold — a decade-by-decade takedown of Silicon Valley's attempts to make face computers happen.

The progression: Google Glass (2013, 'you really want this' → 'no'), Microsoft HoloLens (2016, enterprise → 'maybe but also no'), Meta Ray-Bans (2023, normal + AI → 'maybe? ... actually no'), Apple Vision Pro (2024, $3,499 face helmet → 'absolutely not'), Snap (2026, 'for real this time' → 'we admire the persistence but still no').

289 likes and 32 retweets — a well-executed humor post that also makes a real point about category persistence and market readiness.

The core argument: benchmark grids that show performance as a single number are fundamentally broken in the test-time compute era. GPT-5.5 is much more efficient with its thinking than 5.4, but the benchmark grid doesn't show that because it doesn't control for inference budget.

Safety frameworks gap: responsible scaling policies were designed for the ChatGPT era when test-time compute didn't matter. Today, model capability is a function of how much money you put in — $10 vs $10,000 vs $10,000,000 gives vastly different capabilities. The policies don't address at what budget you should evaluate.

Poker bot progression: early models couldn't do anything. GPT-5.2 made a reverse solver (5x faster than manual). GPT-5.5 basically does the whole solver zero-shot. Noam predicts in 6-12 months, the model will zero-shot his entire PhD thesis.

Erdős conjecture: OpenAI used an internal model to disprove the unit distance conjecture at a very low budget. After announcement, people found 5.5 could also do it — with scaffolding and ~$1,000-$100,000 of compute. Nobody had previously explored what happens when you put $100,000 of compute into 5.5.

On RSI: the models are accelerating lab researchers but can't fully replace them — they lack 'research taste.' Noam doesn't believe in overnight intelligence explosion because test-time compute creates a time bottleneck.

Latent capabilities: there are likely significant capabilities in already-released models that nobody has fully explored, because model release cycles (2-3 months) are shorter than the time needed to push models to their limits.

Multi-agent future: inspired by human civilization (not smarter individuals, but accumulated shared knowledge). Current AI models are 'born into a world, exist for a short context window, and disappear' — eventually they'll coordinate on a large scale.