Overview
Today’s feed had a clear theme: AI teams are chasing better results with smaller, more practical models, while developers get new tools for shipping agents in the places people already work. Alongside that, there were familiar arguments about automation coming for jobs (from banking to surgery), plus a couple of classic internet distractions, from nuclear battery hype to Backrooms nostalgia.
The big picture
The centre of gravity is moving from “biggest model wins” to “useful systems win”. That means better architectures, tighter tooling, clearer safety policies, and distribution through chat platforms rather than shiny new websites. The result is less sci-fi, more production software, with the usual tension between capability, cost, and accountability.
Qwen 3.5 goes after the ‘bigger is better’ idea
Alibaba’s Qwen team rolled out the Qwen 3.5 medium series and the headline is blunt: a 35B model outperforming earlier 235B-era variants on key benchmarks. It is a reminder that training recipes, data quality, and architecture choices can beat brute scale, especially when people want models they can actually run and iterate on.
It also lands neatly in the current appetite for local and hybrid inference, where “fast enough and smart enough” tends to beat “best in the lab”.
Claude Code learns a practical trick: turn websites into tables
Santiago (@svpino) showed a Claude Code skill that pulls live data from websites and returns it as structured output you can use straight away. The tone matters here: less “here’s a long summary”, more “here’s a table you can sort, filter, and export”.
If agents are going to do real work, this kind of grounded, schema-based extraction is what makes them feel dependable rather than chatty.
GPT-5.3-Codex lands in the Responses API
OpenAI pushed GPT-5.3-Codex out to all developers via the Responses API. The subtext is that coding models are becoming less of a novelty and more of an assumed utility layer, especially for agent loops that plan, edit, run, and patch.
For teams already building on OpenAI’s stack, this reads like a nudge to ship, not a research flex.
The price tag reality check for GPT-5.3-Codex
Alongside the launch hype, @scaling01 posted the pricing snapshot: cheap inputs, pricey outputs. That matches what developers feel day-to-day, where the “thinking” part can be affordable, but multi-turn coding assistants can quietly rack up cost when they generate lots of text.
It is not a deal-breaker, but it does push people towards tighter prompts, shorter loops, and better guardrails.
Agents are heading to Slack first, not your homepage
Guillermo Rauch (@rauchg) summed up what many teams are learning the hard way: if you want an agent used daily, put it where work already happens. His point is less about interfaces and more about distribution, adoption, and habit.
The interesting bit is the “build once, deploy everywhere” direction, where the same agent experience shows up across Slack, Discord, Teams, and Google Workspace without rewriting the whole product each time.
Anthropic updates its Responsible Scaling Policy, again
Anthropic published version 3 of its Responsible Scaling Policy, framing it as a response to what worked and what did not since 2023. The gist is a stronger commitment to transparency, with clearer reporting and external review hooks.
It also lands in a tense moment for the industry, where safety language is increasingly tested by politics, contracts, and the urge to deploy faster.
Anthropic rumoured to be pushing deeper into investment banking work
@kimmonismus claimed Anthropic is moving into investment banking workflows, with the caveat that details were not fully confirmed in the thread. Still, it fits a visible trend: high-paid “document and judgement” roles are being productised into repeatable agent tasks.
Whether that turns into job loss or job redesign will depend on who controls the tools, and how quickly firms change what they expect from junior staff.
Musk’s three-year prediction for robot surgeons sets off the usual fight
@r0ck3t23 amplified a clip of Elon Musk claiming Optimus could beat the best human surgeons within three years, “at scale”. The timeline is the spark, but the underlying debate is bigger: where autonomy is acceptable, how liability works, and what “better” even means in medicine.
Robotic assistance is already normal, but full autonomy is a different category, with regulators and hospital procurement moving slower than interview bravado.
Coin-sized nuclear batteries, big claims
@Kekius_Sage posted the eye-catching claim: a coin-sized nuclear battery with 50 years of output. Betavoltaics are real, but the practical story is usually about tiny power budgets and niche use-cases, not replacing your phone charger.
Still, it is easy to see why the idea spreads, it scratches that itch for a future that arrives in a single headline.
The Backrooms is getting the cinema treatment
@levelsio flagged that “The Backrooms” is becoming a real film. It is a neat example of internet-born mythology getting a proper production pipeline, taking a shared vibe and turning it into a ticketed experience.
Whether it works will come down to craft: liminal horror is easy to meme, harder to sustain for a full runtime.
Episode #321: 24 February 2026
