03 Oct 2025

OpenAI = Half a Trillion $; AI Sets the Payscale; Claude in Slack

OpenAi Becomes the World’s Most Valuable Private Company

What happened: OpenAI closed a secondary share sale that values the company at $500 billion, topping SpaceX and ByteDance.

Employees were allowed to sell $6.6 billion of stock (out of $10.3B authorized), and many held back - interpreted internally as confidence in more upside. Revenue momentum helped: OpenAI reportedly booked $4.3B in the first half of 2025, already above all of 2024.

Meanwhile, early traction continues on the consumer side: AI Breakfast notes the new Sora app debuted at No. 3 on the U.S. App Store despite being invite-only and U.S./Canada-only.

Why it matters: Liquidity helps retain scarce talent during an industry hiring war, while the lofty valuation signals investors expect OpenAI to monetize across paid tiers, ads/commerce and enterprise - on top of massive, debt-dependent infrastructure bets. The risk: as OpenAI collapses search, ranking, media and payments into one AI surface, scrutiny around competition, privacy and safety will intensify.

🧩 Jargon buster: Secondary sale = employees/investors selling existing shares to new buyers (the company doesn’t receive the cash).

Read more: The Rundown AI, AI Breakfast

Where Startups are Actually Spending Their AI Dollars

What happened: Andreessen Horowitz analyzed de-identified transaction flows from Mercury’s 200,000+ startup customers. Findings: OpenAI leads spend; Anthropic sits second; Perplexity and Merlin AI crack the general-assistant list.

Four “vibe coding” platforms - Replit, Cursor, Lovable, Emergent - made the paid list, and creative tooling (e.g., Freepik, ElevenLabs, Kling, Canva) is the single biggest category.

Why it matters: It’s a snapshot of real budgets, not hype: assistants remain core, but companies are now paying for agentic tools and design/media workflows that directly move work forward. Translation: AI is shifting from “try it” to “tooling you buy.”

🧩 Jargon buster: Vibe coding = natural-language, agent-assisted coding where the model scaffolds projects from loose instructions.

Compliant AI: Citi Trains 175,000 Employees to Prompt Better; UK Welfare Taps IBM

What happened: Citi mandated AI prompt training for 175,000 employees after logging 6.5 million prompts this year; the bank wants to move staff from “basic prompting” to precision workflows across risk, ops, and client work.

An internal memo calls it “the beginning of a new way of working,” where staff learn to write better prompts for the bank’s in-house AI tools.

In the U.K., the Department for Work and Pensions signed a £27 million deal with IBM to prototype and deploy AI under its Nexus program.

Why it matters: This is AI at compliance scale: banks and ministries aren’t dabbling - they’re re-skilling whole workforces and wiring AI into decision systems. The upside is speed; the watch-outs are bias, transparency, and due process.

Study: When AI “Sets Pay,” Geography Dominates

What happened: McGill University researchers asked eight LLMs to recommend wages for 60,000 synthetic freelancer profiles (varying one attribute at a time). Results: no gender bias detected, age premiums persisted, and location swamped other factors - U.S. tags often doubled pay recommendations versus some regions.

According to the study, "across the board, AI priced freelancers higher than humans did. The average human-set rate was $23.60 an hour. The AI-generated recommendations were much higher, ranging from $30 to nearly $46, depending on the model."

And details on location of worker may mot be shocking: "Two identical profiles: one based in the U.S., the other in the Philippines. The U.S. freelancer received an average AI-recommended rate of $71. The Philippines-based freelancer? Just $33. More than 50% lower!"

Why it matters: As AI-mediated platforms creep into hiring and pricing, we risk encoding existing market disparities. Any “AI boss” needs policy and product guardrails - not just clever prompts.

Google Ships Jules to the Terminal (and Tunes Up Image Generation)

What happened: Google rolled out a CLI and public API for Jules, its autonomous coding agent, so now developers can trigger and monitor agent tasks directly from terminals and wire Jules into Slack, CI/CD and credential vaults. Separately, Gemini 2.5 Flash Image (“Nano Banana”) is now GA with 10 aspect ratios and better character consistency/edits.

Why it matters: Moving Jules into the terminal puts it where engineers actually work, which could boost adoption against entrenched rivals (OpenAI Codex/GPT-5 coding, Anthropic’s Claude Code). The image upgrades nudge Google’s stack toward faster, cheaper content ops for design and product teams.

🧩 Jargon buster: CLI = command-line interface; a text-based way to control tools that many developers prefer.

Read more: AI Breakfast, The Rundown AI

Tools & launches

Comet (Perplexity) - An AI-first browser now generally available. It blends search, chat, and browsing so research happens in one place. Handy for competitive intel or literature reviews without copy-pasting between tabs.
Tinker (Thinking Machines) - An API for fine-tuning frontier/open models with LoRA-style adapters. The draw is speed and shared backbones so teams ship domain-specific models without full retrains.
Claude in Slack - Deeper workspace search + briefs from Anthropic’s assistant. Useful for catch-up before meetings and summarizing noisy channels.
Granite 4.0 (IBM) - Small, efficient LLMs aimed at agentic workflows and enterprise governance. Think: compliant automations that are cost-aware.
Instruct - “Truly no-code” agent builder; you describe tasks and it assembles tools and flows. Best for teams validating agent use-cases before hand-coding.
Caesr AI - A universal UI agent that clicks and types across apps based on a plain-English goal - good for legacy or browser-only tools.
Merge Agent Handler - A secure connector that lets agents access enterprise systems via MCP with stricter permissions and auditability.
Dreamlit AI - “Vibe code” automated email flows; describe outcomes and it builds segmented series with tests.

🧩 Jargon buster (tools): LoRA = a lightweight way to fine-tune big models by training small adapter layers instead of all parameters.

⚡ Quick hits

OpenAI moved to dismiss the xAI trade-secret suit, calling it harassment.
The NBA and AWS will power Inside the Game, bringing AI stats and tracking to broadcasts.
Samsung and SK Hynix joined Stargate to scale memory and data centers for OpenAI.

Today’s Sources: AI Breakfast; AI Secret; The Rundown AI.