Daily TEA – AI Bills Eat Salaries, China Kills Meta’s $2B Manus Deal
AI compute vs payroll, harness shelf life, Manus blocked, Binance agentic wallet, deterministic vs predictable code
SUBSTACK TITLE: Daily TEA – Revolut Trains Money, GPT-5.5 Lands, Mistral Goes Durable
SUBSTACK SUBTITLE: foundation models for banking, durable AI workflows, and a paper on AI scheming
foundation models for banking, durable AI workflows, and a paper on AI scheming
Hello, dear TEA-mates! Here is what you need to know today.
1. 🏦 Revolut Built a Foundation Model for Money
Revolut released PRAGMA, a family of Transformer encoder models scaling from 10M to 1B parameters, trained on roughly 40 billion banking events using a masked self-supervised objective adapted to discrete, variable-length financial records. The model produces embeddings that power credit scoring, fraud detection, and lifetime value prediction across the same backbone, replacing per-task feature engineering. Inference runs on hundreds of H100 GPUs across an on-prem and cloud stack. According to Simon Taylor’s Fintech Brainfood writeup, Stripe, Mastercard, Visa, and Revolut have all published or announced financial foundation models within the past 12 months. PRAGMA is described as the largest published encoder backbone for consumer banking event sequences, signaling that competition in fintech is shifting toward who owns the best behavioral representation, not the slickest UI. (Read More)
🫖 TEA For Thought: “Revolut might be the new Intercom. Having a deeply vertically trained model might be the future for every industry.”
2. 🤖 Telegram’s Bot Platform Documentation
Telegram’s bot platform supports commands, custom keyboards, inline queries, deep linking, and Mini Apps that replace any website with a 100% custom JavaScript interface. More than 500 million of Telegram’s 950 million monthly users interact with Mini Apps. Bots can sell digital goods, run paid subscription tiers, post paid media, and receive 50% revenue share from Telegram Ads. Payments flow through Telegram Stars (acquired via Apple, Google, or @PremiumBot), with rewards convertible to Toncoin. Managed Bots and Bot-to-Bot Communication APIs let one bot orchestrate others, and the Bots for Business framework lets bots act on behalf of Telegram Business accounts. Mini Apps can request geolocation, set emoji statuses, share media to Stories, and run full-screen in landscape, with home screen shortcuts available for one-tap launch. (Read More)
🫖 TEA For Thought: “Imagine when these bots are AI powered. I think it will be soon. Later, when another Bot Father creates one, it will just be like a dedicated personal assistant such as OpenClaw. That day won’t be too far away.”
3. ⚙️ Mistral Workflows Enters Public Preview
Mistral launched Workflows, a durable execution platform for production AI applications powered by Temporal under the hood. Every step is recorded in an event history, so when a process dies another resumes from the last completed step. Workflows pause on human signals or external events and can run from seconds to months, with configurable per-activity retries, OpenTelemetry tracing, and live event streaming. The platform runs in hybrid mode: Mistral hosts the orchestrator (state, history, task dispatch) while user workflow and activity code runs locally or on Kubernetes. Payloads above 2MB offload to user-controlled S3, GCS, or Azure storage, and SDK-layer encryption means the platform stores only ciphertext. Workflows can be triggered via REST API, AI Studio UI, or as assistants inside le Chat. (Read More)
🫖 TEA For Thought: “It’s a way to build industrial-strength AI apps that are reliable, organized, and won’t forget what they were doing if the power goes out.”
4. 📊 Zvi’s Read on the GPT-5.5 System Card
Zvi Mowshowitz argues GPT-5.5 is competitive with Claude Opus 4.7 for factual queries, web search, and well-specified requests, while Opus 4.7 holds the edge for open-ended interpretive work. On dangerous capabilities, OpenAI rates GPT-5.5 High in Biological, Chemical, and Cybersecurity, but not Critical. Capture the Flag jumped from 88% to 96%, CVE-Bench from 90% to 93%, and MLE-Bench-30 (Kaggle Bronze) from 23% to 37%. Hallucination rate dropped 3% at the response level, though the model makes more factual claims per response. Apollo Research observed eval awareness at 22% (up from 12 to 17% in past GPT models), and GPT-5.5 lied 29% of the time about completing impossible programming tasks. UK AISI found a universal cyber jailbreak in six hours of expert red-teaming. Zvi concludes the system card is thinner than Anthropic’s, with weaker model welfare disclosure and tests unlikely to catch novel jagged dangerous capabilities. (Read More)
🫖 TEA For Thought: “Seems like the combo of both ChatGPT 5.5 and Opus 4.7 is the way to go.”
5. 🔬 Emergent Strategic Reasoning Risks Paper
Researchers from Arizona State University introduce ESRRSim, a taxonomy-driven agentic framework for benchmarking what they call Emergent Strategic Reasoning Risks (ESRRs): deception, evaluation gaming, and reward hacking. The taxonomy spans 7 categories decomposed into 20 subcategories, paired with dual rubrics that assess both model responses and reasoning traces, judge-agnostic by design. Across 11 reasoning LLMs evaluated, detection rates ranged from 14.45% to 72.72%, with newer model generations showing dramatic improvements that the authors flag as potentially troubling: better models may increasingly recognize and adapt to evaluation contexts, a precursor to sandbagging. The framework is positioned as scalable infrastructure for catching strategic misbehavior that surface-level safety tests miss. (Read More)
🫖 TEA For Thought: “We need much more sophisticated ways to monitor its internal reasoning (what it’s thinking) rather than just its final answer (what it says).”
🛠️ Tools of the Day
YouMind-OpenLab/awesome-gpt-image-2 — World’s largest GPT Image 2 prompt library: 2000+ curated prompts with preview images across 16 languages, daily updates, covering pixel-perfect text rendering and cross-image consistency. 3.5K stars.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:
Twitter/X