A winter spent by The Pond
A scary, angry, sweet winter.
Stand firm against authoritarianism
Act now or forever bear the dismay of history’s gaze.
Last October, I warned: “Accept the reality of the risk cast by the future. Act now to prepare.”
ICE has become a lawless, unaccountable agency—a starving, rabid dog let off its leash. Generally, rabid dogs are not empowered by the full force of the US federal government. This one is. That same government has declared its political opponents to be “domestic extremists” and pursues a naked agenda of persecution.
We have reached the “street execution of innocent civilians” phase.
Even a few weeks later, I still think about Alex Pretti. He died defending two people from masked men harassing them with pepper spray.
By the accounts I have read, Pretti lived his life as a generous, kind man.
A while back, I sent some friends the following message:
I keep looking at Alex Pretti’s photo and crying. What an absolute hero. Exactly the kind of man I want to be. Spent his last moments protecting two women from ICE thugs, a true gentleman
I will not honor Alex Pretti through words. He deserves better than that. I will honor him through the way I live my life—through effective action, brave action, necessary action.
Many of my readers are in tech. Quite a few tech companies enable ICE. You want to help? Do you actually want to help? Then help dismantle a piece of that supply chain. Take an honest, no-rationalization look at your company’s part.
Obvious enablers include Palantir, Amazon, and Microsoft. Ask your local AI for some deep research. Then, consider the organizational structure in your company. Have high-level executives already spoken out? Reach out to them internally—ask what they think and if there’s any way you can help. Find creative pressure points and push.
Speaking out puts one at some risk, yes. Even though I’m doing nothing wrong by the laws of this country, we aren’t operating by laws anymore. Maybe it’ll be inconvenience—I might find my Global Entry to be revoked the next time I go overseas. Or maybe I’ll have trouble getting back into the country. Or maybe, should America continue her descent—maybe I’ll just be snatched as an “enemy of the state” just for speaking my mind.
But who would I be, to tremble at such a safe distance from Minnesota—while all is soft and floral in peaceful Berkeley—while I sit upon privilege and influence? Compared to the heroic protestors and rapid response operators on Signal, my risk is tiny.
In any case, my momma didn’t raise a coward.
Fuck ICE.
Look, guys, I don’t have a smooth transition here. Sadly, this ICE secret police stuff is part of our lives now. But also, there’s more to our lives than that. And I want to talk about that other stuff, too.
New writing
No instrumental convergence without AI psychology. Instrumental and success-conditioned convergence both require AI psychology assumptions, so neither is just a “fact about reality.”
Consistency Training Helps Stop Sycophancy and Jailbreaks. Simple & effective: train the AI to behave as if the jailbreak were not present. Explores activation-level training of Gemini 2.5 Flash.
Output Supervision Can Obfuscate the CoT. We challenge the assumption that output supervision preserves Chain-of-Thought faithfulness. Instead, reinforcing final outputs warps the CoTs.
2025-era “reward hacking” does not show that reward is the optimization target. “Reward hacking” is usually specification gaming, not reward signal optimization. My 2022 post stands.
Recontextualization Mitigates Specification Gaming. Resist specification gaming by generating data with anti-misbehavior prompts, then training on pro-misbehavior prompts.
Privacy guide updates
Since the last newsletter, I’ve continued refining my privacy guides based on reader feedback. An opinionated guide to privacy despite authoritarianism and Advanced privacy despite authoritarianism now include:
New section on downloading Android apps via Obtainium (downloads directly from developers). F-Droid deprecated.
I realized that macOS shares iOS’s VPN leak problem, sabotaging privacy for those with harsh threat models! The VPN leak warning now explicitly covers macOS, not just iOS. The only reliable workaround is either ditching iOS/macOS or connecting through a router that enforces VPN protection for all traffic. I wrote a new subsection with steps.
Open source
I built alt-text-llm, a Python package for AI-powered alt text generation. It scans markdown files for images and videos missing meaningful alt text, generates context-aware suggestions using the LLM of your choice, and lets you interactively review and apply them.
Team Shard recruitment page
Apply for Alignment Mentorship From TurnTrout and Alex Cloud: Team Shard consistently graduates skilled researchers, does good work, and yields three months of wholesome growth.
Through the MATS program, Alex Cloud and I help alignment researchers grow from seeds into majestic trees. Applications for January 2027 will likely open around the summertime.
I took the 10% pledge
For the rest of my life, I’ll donate at least 10% of my post-tax income to effective charities.
Site Updates
520 commits since the last newsletter, including:
Search match fading. When you navigate to a page from search, matches glow green then gracefully fade over 5 seconds.
Dynamic site statistics. Commit counts and other metrics now auto-populate:
Build checks ensure paragraphs don’t end without proper punctuation.
Card preview images are now checked to be under 300KB, ensuring compatibility on platforms like Signal.
OpenTimestamps reliability improvements. What a tragedy would befall my intellectual legacy if my commits were not cryptographically verifiable!