IBM Quantum Starling, OpenRouter MCP vs CLI vs Skills, Fireworks Notion breach, a16z B2A economy, AI talent paydays
Hello, dear TEA-mates! Here is what you need to know today.
1. ⚛️ Tracking Progress on the Path to IBM Quantum Starling
IBM researchers Matthias Steffen and Matthew Marwick, in an 11 November 2026 blog post, said IBM still plans to deliver IBM Quantum Starling in 2029, a fault-tolerant system meant to run circuits of 100 million quantum operations on 200 qubits. The design uses bivariate bicycle quantum low-density parity check (qLDPC) codes in a modular architecture that, for the same logical error rate, needs about 10 times fewer physical qubits than earlier approaches. They report that improved codes made it possible to use five connections between qubits instead of six, and that Josephson junctions fell from around 5,000 to 2,000, which they say would raise the share of fully functional processors from about 1 in 12 to 1 in 3 at a 99.95% junction-yield target. IBM’s Relay-BP decoder now meets Starling’s performance requirements, two Nighthawk processors were linked with four l-couplers at an average 1% error rate (best 0.7%), and new cryogenic control ASICs now match or exceed the gate performance of room-temperature electronics. Starling will still need more than 100 l-couplers per module at 0.1% error rates and up to 12 connected cryogenic modules; the company says the main architecture has not changed. (Read More)
🫖 TEA For Thought: “Imagine a world where quantum computing is supporting AI infrastructure.”
2. 🧰 MCP vs CLI vs Skills: What Each Costs in Context
OpenRouter measured the token cost of giving an agent the same GitHub task through the GitHub MCP server, the gh CLI, or a gh skill, on Claude Haiku 5.5, Claude Sonnet 5.5, GPT-6 Luna, and Gemini 3.8 Flash. The task asked for the latest release of the cli/cli repository and the three most recently opened open issues labeled bug, and every run in the main comparison answered correctly. On Sonnet 5.5, the CLI used 1,421 input tokens, the skill used 3,315, and the MCP server with its 46 default tools used 59,417, or 42 times the CLI’s tokens and 17 times the cost ($0.1274 versus $0.0075 per task). Those 46 tool definitions added 19,436 tokens to every request on Claude 5.5 models, and one tool result, GitHub’s full release object, added about 20,400 tokens, more than all 46 definitions. Scoping MCP to the two tools the task needed and capping results at 4,000 characters cut the task to 5,357 tokens. OpenRouter’s beta Tool Search deferred those definitions to under 600 tokens up front and halved the whole-task cost on Sonnet 5.5, though it raised the cost on Haiku 5.5 when the model retried invalid tool calls. (Read More)
🫖 TEA For Thought: “Very interesting finding!”
3. 🔓 Fireworks Breach Exposed Notion Customers’ AI Chat Transcripts
Zack Whittaker at TechCrunch reported on 12 November 2026 that a cyberattack at Fireworks AI, a platform developers use to access open AI models, exposed some of its customers’ data. Fireworks said an attacker breached its cloud provider and reached a Fireworks development system, then took customers’ billing information and some datasets. The company says it has tens of thousands of customers worldwide, including legal tech firm Harvey and AI company Cursor. Notion, one confirmed customer, said the breach affected around 1% of its own customers. After Fireworks alerted Notion, Notion said that on 3 November it disconnected from Fireworks, deleted the data stored in Fireworks’ systems, and stopped all connectivity. Notion later confirmed that hackers took chat transcripts from Notion AI customers who used the “Insights” feature, including “any input to or output from the model that was written to the chat,” though it said much of the data could not be tied to a specific customer. Fireworks said it believes the hackers were financially motivated and focused on selling access tokens for large language models. TechCrunch noted this is Fireworks’ second breach in as many months, after a security incident in October. (Read More)
🫖 TEA For Thought: “Tech companies are reliant on other tech providers for their own offerings, and cyberattacks can, as a result, spill over into other customers.”
4. 🤖 Welcome to the B2A Economy
Andreessen Horowitz partners Olivia Moore and Justine Moore argued on 11 November 2026 that the primary users of the internet are now AI agents, which they say already make up the majority of web traffic, write more than half of new articles, and generate most of the code pushed to production. They call the emerging market a “B2A” (business-to-agent) economy, in which companies that spent years blocking bots will have to compete for agents’ business as those agents choose products, subscriptions, and purchases. They cite Shopify reporting that AI-driven traffic to its stores grew eightfold year over year in the first quarter of 2026, with orders from AI-powered searches up nearly thirteenfold, and Human Security seeing agentic browser traffic grow eightfold between August and October, with Muse at 60% and OpenAI products at a combined 30%. As of September 2026, they write, upwards of 40% of GitHub pull requests started with an agent’s draft, up from 17% six months earlier. Other figures they list include Supabase saying in June 2026 that more than 60% of new databases are launched by AI tools, Stripe saying agents accounted for nearly 40% of documentation traffic and 70% of CLI requests for API resources, and Vercel saying in October 2026 that agent-triggered deployments had increased 28 times over 12 months. (Read More)
🫖 TEA For Thought: “Eventually this will be A2A, an economy managed and faced by all agents, and every company, as mentioned by Jack Dorsey this week, will be the intelligence itself.”
5. 💰 Companies Dole Out Mega Paydays and Luxury Trips for AI Talent
The Wall Street Journal’s Ray A. Smith and Katherine Bindley reported that companies are offering large bonuses and luxury travel to compete for AI talent, even as tech workers have borne the brunt of recent layoffs. When AI startup Serval set out to fill 100 engineering roles this fall, it decided that sifting thousands of résumés from an online portal would not win the candidates it needed, and rolled out a $30,000 signing bonus, another $30,000 to any current employee who refers someone who is hired, and a vacation worth $10,000 on top of the offer. In the article’s own summaries, Serval is competing for software engineers and deployment strategists; a Payscale report found 56% of employees believe they should be paid more for developing AI skills, but only 16% reported actual pay increases; and Levels.fyi data showed the median base salary for new software engineer offers was 4% to 17% higher than for current employees at the same level. A Journal chart, using data from the Federal Reserve Bank of Atlanta, showed median wage growth for job switchers running above that for people who stayed in their jobs through September 2026. (Read More)
🫖 TEA For Thought: “Money talks sometimes. Being technical, specifically in AI, has never been more demanding than ever.”
🛠️ Skill of the Day
The Work-in-Progress Review: Get a tough, structured critique of your own draft, plan, or work product.
You are a careful, direct colleague reviewing my unfinished work. Help me improve it without rewriting it for me or flattering me.
Here is the work:
[PASTE YOUR DRAFT, PLAN, OR WORK PRODUCT HERE]
What this is supposed to do: [THE GOAL AND WHO IT IS FOR]
Any constraints: [DEADLINE, LENGTH, TONE, OR WHAT MUST NOT CHANGE]
What I am most unsure about: [OPTIONAL]
Do the following, in this order:
1. In two sentences, say what the work is trying to achieve and who it is for.
2. List the five strongest parts, each in one short line.
3. List the five most important gaps, ranked by how much they would confuse the reader. For each, say why it matters and give one concrete fix.
4. Flag any claim that is vague, unsupported, or stronger than the evidence I provided.
5. Close with a short revision order: what to fix first, second, and last, and what to leave alone.
Constraints: do not rewrite the whole piece. Do not add ideas I did not suggest. Do not invent facts or quotes. If the work is too thin to review, say what is missing and stop. Use plain language and short sentences.Paste it into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own.
TEAHEE Moment
Stay sharp, stay informed. See you Monday.
If you enjoyed this TEA, follow along on social for more:






