I keep a messy bookmark folder, and once in a while I look back at a week of it and notice the items are arguing with each other. This week gave me five. A venture thesis about who owns your AI memory. A Google forecasting model being sold as a stealth weapon. A grey market in China reselling Claude access. A thousand excess deaths in France. And a viral line claiming Chinese intelligence admitted aliens walk among us.
They look unrelated. They are not. Each is a claim about control, prediction, or access, and each rewards the same discipline: ignore the volume, look at who benefits from the framing, and name the actual mechanism underneath. So let me do that.
Who owns your memory
Nick Grossman at Union Square Ventures published a piece called "The rebel alliance," a sequel to an earlier one titled "You don't own your memory." The argument is sharp. Frontier models are bumping into a scaling ceiling, and the next gains come from adapting to your personal context. The one dataset the big labs do not have is your accumulated memory, scattered today across closed tools that barely sync. There is a fork ahead. Either a few large companies absorb that memory into closed systems and promise not to be evil, or a distributed ecosystem lets you carry your memory across competing AIs that have to earn your trust.
The good idea here, borrowed from Alex Komoroske, is the move from "won't be evil" to a system that is structurally unable to be evil, because power is distributed rather than concentrated. The plumbing for this is already taking shape. The Model Context Protocol, pushed by Anthropic, is an open standard that lets an AI connect to your calendar, email, and files, often described as USB-C for AI. The financial-data company Plaid is the obvious precedent for a memory infrastructure layer.
I find the structural argument persuasive. I also note plainly what it is: a venture capitalist talking up a category his fund invests in. Data-portability products have a long history of losing to convenience, and "users will switch because it works better" is an assertion, not a result. There are real open problems too, chiefly that no mature mechanism exists for handing an agent scoped, auditable authority to act on your behalf, and that persistent cross-service memory breaks privacy frameworks built for siloed, local data. Read the thesis as a thesis, not as prophecy.
A forecasting model wearing a disguise
The second bookmark, in Spanish, announced that Google had "silently released" an AI that predicts patterns: sales, market prices, web traffic, energy demand, crypto volatility. It is called TimesFM, trained on 100 billion real data points, zero-shot, runs locally. Probably the most something, the post cuts off.
The model is real and genuinely good. The "silent" framing is nonsense.
TimesFM was posted as an arXiv preprint in October 2023, accepted at ICML 2024, and open-sourced under Apache 2.0 in mid-2024. It is a 200M-parameter decoder-only model that treats stretches of time as patches, the way a language model treats tokens, and it forecasts unseen series without per-dataset training. The version arc is the interesting part. After scaling to 500M parameters in 2.0, Google went back down to 200M in 2.5 and still took the top spot on the GIFT-Eval benchmark, because stretching context from 512 to 16,000 points mattered more than adding parameters. It is wired into BigQuery ML, Sheets, and Vertex. You can pip install it.
The one claim worth pushing back on is the one the post leans on hardest. Financial and crypto prediction is exactly where these models are weakest. Return data has a low signal-to-noise ratio, fat tails, and structural breaks, and independent work found that applying TimesFM to prices directly gives unsatisfactory results; even after fine-tuning on financial data, crypto performance stayed poor. Amazon's Chronos-2 also handles multivariate forecasting, which early TimesFM did not. The model is excellent for cold-start forecasting across many ordinary streams. It is not a money printer, and dressing a published academic model as a leaked secret is just a way to harvest attention.
Block a flow and it reroutes
The third bookmark is a line from Zilan Qian's ChinaTalk investigation: "History teaches us that access blockage rarely stops determined users." The piece documents China's grey market in what developers there call transfer stations, API proxies that sit between a developer and Anthropic, forwarding requests from an allowed location and taking payment in RMB over WeChat or Alipay. They run openly on GitHub, Taobao, and Telegram, and at peak they sell a dollar of Claude tokens for one yuan.
How they stay that cheap is a mix of the clever and the criminal:
- bulk-registering free developer accounts to farm the $5 API credits
- splitting one $200 Max subscription across dozens of users
- using stolen card details to open accounts at zero cost
- paying real people in lower-income countries to clear the photo ID and live selfie checks
The real business is not the markup. The operators log every prompt and response that passes through, and for coding agents that means full reasoning chains and human-verified outputs. Several developers told Qian the access is loss-leader customer acquisition and the harvested logs are the product, valuable for distilling cheaper models. Anthropic restricted Chinese-controlled entities in September 2025, a move Dario Amodei said cost the company hundreds of millions in revenue, and each new verification layer has spawned a matching evasion market rather than shrinking access. This is the Great Firewall lesson in miniature, and it loops back to Grossman's point from the other direction. When you geoblock a flow, you do not end it. You push it onto infrastructure that is harder to see and harder to govern.
The story that actually matters
Then there is France. Santé publique France reported around 1,000 excess deaths during the late-June 2026 heatwave, and said the true figure would be higher. That number deserves to be stated carefully, because careful is how you respect it. It is a preliminary excess-mortality estimate drawn from digital death certificates, which capture only about 60 percent of deaths nationally, not a confirmed count of people certified as dying of heat. Daily deaths peaked above 1,400 on the Thursday and Friday, against 900 to 1,000 in April and May. Eighty-five percent of the dead were 65 or older, and the sharpest rise was among people dying at home, concentrated in the Paris region.
France recorded its hottest day ever on 23 June, with Météo France putting 54 departments under red alert. Germany kept resetting its national record. Across Europe the WHO counted more than 1,300 excess deaths since 21 June and Tedros called heat a silent killer. The mechanism is not mysterious. A heat dome sat over a baseline warmed about 1.4C since pre-industrial times. World Weather Attribution found the event would have been virtually impossible without human-caused warming, and that the dangerous overnight temperatures were roughly 100 times more likely than two decades ago.
In the replies people blamed data centers and concrete. France does host a large data-center sector, somewhere above 320 facilities, and urban heat islands are real, but both are local amplifiers, not the cause of a continental heat dome. The deflection is worth naming because it does the same work the TimesFM hype does, just inverted: it points your attention away from the boring, documented mechanism. Of these five bookmarks this is the one with the highest stakes and, predictably, the least spectacle. People dying alone in apartments in Île-de-France do not trend.
The confession that never happened
The last one announced that "the Chinese CIA admit extraterrestrials live among us right now." No source, no link. It is false, and it is a stitched-together mashup, so I will say so plainly before I say anything more interesting.
There is no record of China's Ministry of State Security, the agency people mean by Chinese CIA, saying any such thing. The MSS did start public messaging in 2023 with a WeChat account, all of it about counter-espionage, which is probably what makes a fake "MSS admits" headline feel plausible. The "aliens among us" line traces to Sun Shili, a retired Foreign Ministry translator who led a civilian UFO hobby club and appeared on Ancient Aliens around 2010. A civilian enthusiast's opinion is not a state position. Separately, in 2022 China's FAST radio telescope was reported to have caught suspicious narrow-band signals, and a SETI collaborator on the project, Dan Werthimer, identified them as terrestrial radio interference. Three unrelated facts, fused and inflated into a confession.
That is the honest read of this particular post. But I am not going to use it to wave the whole subject away, because the interesting detail is the one that keeps repeating. The FAST signal story was published by state-affiliated media and then quietly deleted. Publish, then pull. That reflex is the part I would not file under nothing-to-see-here. The convenient official line, terrestrial interference, case closed, is exactly the kind of tidy explanation that earns scrutiny rather than relief, and the disclosure story over the past few years has given serious people reason to keep asking. So I hold two things at once without strain. This specific headline is fabricated. The larger question of what we are and are not being told is open, and the deletions are evidence of an instinct, not an answer.
The habit
Pull the five together and the shape is the same each time. A VC sells a structure as inevitability. A lab's open model gets repackaged as a secret weapon. A geoblock manufactures the market that defeats it. A heat dome gets blamed on data centers. A deleted radio signal becomes an alien confession. In every case the move that works is the same. Discount the loudness. Ask who profits from the framing. Find the number, the date, the source, the actual mechanism. And then notice which story got buried under the rest, because that is usually the one that was costing real lives.