Daily TEA – Amazon Ships AWS for Everything Else
ASCS, fine-tuning safety drift, AWS hires 11K interns, memory vs lookup, the stablecoin moment
Hello, dear TEA-mates! Here is what you need to know today.
1. 📦 Amazon Turns Its Shipping Network Into Another AWS
Amazon launched Amazon Supply Chain Services (ASCS), opening its global fulfillment network to outside companies and putting it in direct competition with DHL, UPS, and FedEx. Initial customers include Procter & Gamble, 3M, Lands’ End, and American Eagle Outfitters, spanning automotive, healthcare, electronics, apparel, and food. ASCS bundles freight, distribution, fulfillment, and parcel shipping, letting partners store inventory in Amazon’s fulfillment centers and use its trucks, aircraft, and delivery vehicles. The move extends a 2023 pilot that shipped products from factories on behalf of third parties. Peter Larsen, VP of ASCS, says the goal is to give “any other business access to the same cost efficiency, reliability, and speed” Amazon built for itself. The blueprint mirrors AWS, which started selling spare web infrastructure to outsiders in 2006. (Read More)
🫖 TEA For Thought: “Amazon has been doing this all along. Whatever it uses internally can be packaged and sold as a service. It’s very, very brilliant.”
2. 🎛️ Fine-Tuning Foundation Models Quietly Breaks Their Safety Guardrails
A new report from the Center for Democracy & Technology (CDT) and MIT, “Out of Tune: Fine-Tuning Foundation Models Leads to Unpredictable Safety Drift,” shows that even benign domain-specific fine-tuning can degrade a foundation model’s safety behavior in unpredictable ways. Empirical tests on medical and legal models found that minor modifications, regardless of model size or compute budget, caused significant “safety drift,” producing harmful content or dangerous advice the base model would have refused. The authors argue this undermines governance frameworks like the EU AI Act that lean on compute thresholds as risk proxies. Instead, safety should be treated as an emergent, fragile property requiring continuous evaluation across the full supply chain. The report calls for a shared-responsibility model: upstream developers harden base models, downstream deployers run rigorous domain-specific safety tests. (Read More)
🫖 TEA For Thought: “One has to know that scientists, even the top researchers who invented these LLMs, do not understand how they work. It is always predicting the next token, and that prediction is never the same because of the neural network. Chaos is the norm. Order is the outlier.”
3. 👶 AWS CEO Says Software Engineering Demand Is “Accelerating,” Plans 11,000 Interns
At Amazon’s What’s Next with AWS event, AWS CEO Matt Garman said Amazon is hiring “just as many software developers as we ever had” and sees demand “really accelerating.” The company plans to bring on 11,000 software engineering interns globally in 2026, in line with prior years. Garman pushed back on takes from Anthropic’s Boris Cherny and Andreessen Horowitz’s Martin Casado that AI coding tools could erase the role, calling the idea of replacing junior employees with AI “one of the dumbest things I’ve ever heard.” He said the job is shifting: authoring Java snippets matters less, while building applications and solving customer problems matters more. The hiring push lands alongside roughly 16,000 corporate layoffs earlier in 2026, which Amazon says were not primarily AI-driven. (Read More)
🫖 TEA For Thought: “Big techs are laying off mid-managers while hiring interns. Young people with new ways of using and thinking with AI tools are the ones being brought in to build.”
4. 🧠 New arXiv Paper Argues Agent “Memory” Is Actually Just Lookup
In “Contextual Agentic Memory is a Memo, Not True Memory,” authors Binyan Xu, Xilin Dai, and Kehuan Zhang argue that vector stores, RAG, scratchpads, and context-window management do not implement memory at all. They implement lookup. Retrieval generalizes by similarity to stored cases, while weight-based memory generalizes by applying abstract rules to never-seen inputs. Conflating the two, the paper says, produces agents that pile up notes without developing expertise, hit a provable generalization ceiling on compositionally novel tasks (no amount of context size or retrieval quality fixes this), and become structurally vulnerable to persistent memory poisoning where injected content propagates across all future sessions. Drawing on Complementary Learning Systems theory, the authors note biological intelligence pairs fast hippocampal exemplar storage with slow neocortical weight consolidation, and current agents only implement the first half. (Read More)
🫖 TEA For Thought: “Plato said human knowledge is memory, and to learn knowledge is to recall it from embedded memory. Another reminder that what is being built is Artificial Intelligence after all.”
5. 💵 The Stablecoin Moment: Rails for the Agent Economy
zauth’s “The Stablecoin Moment” tracks how stablecoins moved from speculative crypto to the primary payment layer for the emerging agent economy. The market cap has hit $325 billion, with 2025 transaction volume reaching $46 trillion, surpassing Visa’s 2024 throughput. The GENIUS Act (July 2025) classified stablecoins as payment instruments, letting banks hold and settle in them and clearing the path for institutional adoption. Four protocol standards are racing for the standard slot: x402 (Coinbase) for HTTP-native machine-to-machine micropayments, ACP (Stripe + OpenAI) for merchant checkouts, AP2 (Google) for enterprise authorization, and MPP (Stripe + Tempo) bridging crypto and fiat. The “endgame” thesis (per Mert Mumtaz) is invisible infrastructure: users see “USD” while the backend routes through Solana, Base, or whichever chain is fastest and cheapest. (Read More)
🫖 TEA For Thought: “Which protocol is going to win? We might see it sooner than we thought.”
🛠️ Tools of the Day
docusealco/docuseal — Open-source DocuSign alternative for creating, filling, and signing digital documents, self-hostable with form builder and audit log. 13.2K stars.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:
Twitter/X