Happy Friday, friends.

Two of the biggest labs changed what an answer looks like this week: ChatGPT now replies with buttons and charts, and Claude builds dashboards that show you the query behind every number. In the same days, the mayor of Tirana addressed his city through an avatar from a prison cell, and I tried to sort what is confirmed from what the headlines claimed.

If you write software, the Ping Labs compiler piece comes with a price tag worth stealing. The Arena and Siena rounds show where investors are sending cheques, and the UK regulator story has a date for your calendar: 20 November. Six stories, five minutes.

OpenAI started rolling out GPT-6 with Intelligent UI on 7 October, with Free and Go users getting it a day later. Replies can now arrive as tappable buttons, forms, editable charts and calculators, built from a library of native components that a compiler streams while the model is still writing. TechCrunch's demos included a wing-lift explainer, a hiking map and a savings calculator. OpenAI's page says nothing about developer access, so for now this is a ChatGPT feature, not a platform.

On 8 October Anthropic put Dashboards and Motion into beta. Dashboards connects to Snowflake, BigQuery, Redshift, Databricks, ClickHouse and Salesforce, builds from a plain-language question, stamps each chart with its last refresh and lets you open the query behind every number. Motion writes code to animate your charts and text, exports MP4 and generates no footage. Per Reuters, Dashboards sits on paid plans, Motion on Team and Enterprise, and Docs, Slides and Design left beta for every plan. Anthropic has published no accuracy results for either tool.

OpenAI drew the screen for the consumer, Anthropic drew it for the analyst and put the receipt on the face of it. If you sell dashboards or thin front-ends, the screen stopped being your product this week. Ask Claude Dashboards one real question from your own data, then compare the query it shows with the one your analyst would write.

You’ve put in the work to build an audience. But when the only way to reach your fans is through a social feed determined by algorithms, the platform decides who sees what you make next.

beehiiv gives you the tools to reach your audience directly. On their all-in-one platform, you can turn followers into subscribers, grow your newsletter, and connect it all to your community or podcast. Stop chasing virality on platforms you can’t control. Build a stronger, more sustainable business on beehiiv.

Get started for free.

On 7 October Erion Veliaj, the mayor of Tirana, posted a four-minute Instagram video in which an AI avatar of him stands before the Albanian and city flags. Instagram labelled it as AI-generated. In it he says, as TNW quotes him, "For 20 months my hostage-taking has continued, and I cannot send this greeting except through artificial intelligence." He renews a promise of free school lunches. Veliaj has been in pre-trial detention since February 2025 over corruption, money laundering and asset-concealment allegations brought by SPAK, Albania's special anti-corruption prosecutor. He denies them, and nothing has been tested at trial.

"Running the city from jail with AI" is the viral version, and the evidence supports less: one video. The reports I read don't say who built the avatar, and IBTimes notes that day-to-day duties went to deputy mayors before his detention. He is still the legal mayor after the Constitutional Court ruled his dismissal unconstitutional, in a country where the government already presents an AI "minister" called Diella for procurement.

Watch the label, because it is the only disclosure there is. TNW points out that the EU's AI labelling rules, in force since August, don't cover Albania. Anyone in the Balkans building avatar or voice products should write down this month which words they will refuse to put in a generated mouth, before a client with a good reason asks.

The pingdotgg team published tsc-rs, a Rust port of the Go-based TypeScript compiler, and its README says all 181,711 ported tests pass. The Claude Opus run cost about $24,047 in API spend over two weeks. An earlier attempt with OpenAI models burned over $400,000 in tokens and stalled near 84% compatibility. On the T3 Code project it checks in 7.25 seconds, against 16.10 for tsc 7 on an Apple M4 Pro.

Read the caveats before you install it. The author calls it an early release and says nobody has read a line of the code. The 100% compatibility claim is theirs, and known issues include monorepo errors, stale output with tsc -b and slow memory growth in long editor sessions. It runs on Linux x64 and macOS arm64 only, and every figure is self-reported.

The speed isn't the lesson. The 181,711 existing tests were the acceptance criterion, which made the rewrite something you could buy by the token. Count the tests around the system your team keeps calling too risky to touch. That number tells you whether the rewrite is a project or a wish.

Siena, a New York company founded by Romanians Andrei Negrău and Lisa Popovici, raised a $17M Series A led by York IE, bringing total funding to $29.7M. It sells AI agents for consumer brands (support, in-chat shopping, social replies) to names like SPANX, FIGS and HexClad. The company says agents handle up to 80% of interactions and that hundreds of brands use it, both unverified figures. A voice agent is still in development.

The pricing is the part to copy. Brands pay for results and data tools rather than per seat, so Siena earns only when its agent finishes the job. It competes with Gorgias, Intercom's Fin and Sierra, which means the whole category is converging on that model. The founders are Romanian and the company is headquartered in New York, the same route many CEE founders take: build where the buyers are.

Here is a question for this week: can you describe "done" for your product in one sentence a customer would sign? If you can't, you can't price by outcome, and your competitors may soon.

Arena, the platform formerly known as LMArena, raised $200M at a $3.1B valuation, led by Lightspeed and Khosla, ten months after a $150M round at $1.7B. Company figures put annualised revenue near $30M in January and $100M in June. The money comes from AI Evaluations, which sells performance data from its public leaderboard to labs and enterprises.

With the round came a new alignment leaderboard that ranks models on unauthorised action, false attribution and "deceptive completion", meaning a model claiming work it didn't do. The results are preliminary and the methodology isn't public yet. OpenAI models hold the top spots, with Claude Opus 5.5 sixth and Claude Fable ninth. Arena's own line: "static benchmarks break down once models recognize they're being tested."

Someone gets paid whenever output can't be trusted on sight, and this quarter it's the scorekeeper. If you sell AI into companies, expect buyers to ask for an independent score. Collect yours before they ask, because a third-party number lands differently from your own deck.

On 8 October the UK's Information Commissioner's Office said ten developers have changed or committed to change how they handle personal data in training: Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. The commitments cover clearer explanations, easier ways to exercise data rights and tougher safeguard assessments. xAI is missing because the ICO paused engagement and is investigating Grok separately. Unresolved: personal data embedded in trained models, removal after training, and extraction risk.

The second half matters more for builders. The ICO opened a call for evidence on agentic AI, with responses due 20 November, feeding its forthcoming statutory code on AI and automated decision-making. Its director of technology regulation, Richard Nevinson, put it plainly: "the fact AI agents act with autonomy is not an excuse for poor compliance."

This is UK law, not EU, but UK GDPR sits close to the EU version, so read it as the leading indicator. Anyone shipping agents that touch personal data can respond. Start by logging what your agent did with each person's data, task by task. That log is the answer every regulator will ask for.

Marketing: OpenSEO. An open-source alternative to Semrush and Ahrefs that self-hosts via Docker or on Cloudflare's free plan and ships an MCP server for Claude Code. You bring your own DataForSEO key and pay per request, so cost follows usage. Run one keyword gap check on your own site before you renew an SEO subscription.

Productivity: IrisGo for Solopreneurs. A desktop app whose "Watch & Learn" mode records a task you do once and replays it as a workflow, plus daily briefs and email triage. Free during beta, proprietary, and it needs OS-level permissions. I found no independent reviews, so try it on a throwaway account first.

Design: Nano Banana 2.1. Google's new image model takes up to 14 reference images and is available in AI Studio and the Gemini API as gemini-nano-banana-2.1. The docs list no function calling or structured outputs, so it won't slot into agent pipelines yet. Test it on one product shot with several reference images.

Dev: Anthropic's OSS Scanner. Maintainers of critical open-source projects can enrol by pull request to anthropics/oss-scanner for free, model-run vulnerability scans. Eligibility is case by case, and Anthropic warns the reports are model-generated and may skip human review. If you maintain a package other companies depend on, open the PR.

Next edition soon,
Çelik