Gordon
banner
gordon.bsky.social
Gordon
@gordon.bsky.social
Everything around me was someone’s lifework.
Pinned
self-certifying data… over http
Reposted by Gordon
Otoh frequentism is no use at all in a minefield, unless you have enough legs to approximate a normal distribution.
It is difficult to practice Bayesianism in a minefield.
November 28, 2025 at 1:20 PM
big thorny refactors are so much easier with claude code
November 28, 2025 at 12:28 PM
It is difficult to practice Bayesianism in a minefield.
November 28, 2025 at 6:16 AM
"100 of us go to a casino and gamble. If the 28th person is ruined, this has no impact on the 29th gambler. We can compute the casino's return by taking the returns of the 100 people who gambled. If we do this two or three times, we get a good estimate of what the casino's edge is..."
November 27, 2025 at 12:17 PM
Reposted by Gordon
Fairy driven development
November 27, 2025 at 5:16 AM
Reposted by Gordon
Followup to yesterday's post: I'm starting to think as agents and LLM APIs of being a state synchronization problem and that we might look into what the local first folks are doing. Dumped my thoughts here: lucumr.pocoo.org/2025/11/22/l...
LLM APIs are a Synchronization Problem
Maybe the LLM message APIs should be rethought as a synchronization problem.
lucumr.pocoo.org
November 22, 2025 at 12:15 PM
Reposted by Gordon
Plinko PIR tutorial
vitalik.eth.limo
November 25, 2025 at 11:58 PM
Reposted by Gordon
I regret spending years trying to build this rather than refute it. Long-lived identities should never have been an objective to begin with.
November 26, 2025 at 4:12 PM
memento mori
Yeah, which highlights how those aren't unsolved problems just with AIs.

And it's the same for the process of continual learning. I just read someone state that brains are the "gold standard" in that. It seems to be so easy to forget that catastrophic forgetting is very much a feature for us.
November 26, 2025 at 4:17 AM
Reposted by Gordon
Once the user releases their Spirit Bomb, it's so over
this gif, except substitute "language" for "user" and "human intelligence" for "engineer".
November 25, 2025 at 2:34 PM
Realizing that bash/linux have been immortalized in latent space AND we're about to see a bunch of great new Linux sandboxing tools because of this. Yes... ha ha ha... yes!! bsky.app/profile/timk...
Anthropic is giving away all their secrets
November 25, 2025 at 11:37 AM
this gif, except substitute "language" for "user" and "human intelligence" for "engineer".
November 25, 2025 at 10:57 AM
Interop is not a feature. It’s an ecological condition.
The point being, while things like SQLite and CRDTs are not ubiquitous now, it's also not an impossibility that they can be ubiquitous in the future. Ubiquity is not a special property, it's an infrastructural investment.
November 25, 2025 at 6:32 AM
turning a big dial taht says "evil vector" on it and constantly looking back at the audience for approval
November 24, 2025 at 7:34 AM
evil as attractor in behavioral space
i'm going to sound sarcastic here but genuinely one of the most interesting aspects of LLMs is that they are proving that evil is real, by which i mean if you train them to do one bad thing they start doing other bad things that you have not trained them to do
At least once a month, Anthropic puts out an alignment paper d AI behavior where, if you saw it in a science fiction film, you‘d be screaming at the idiot scientists onscreen to stop development

www.anthropic.com/research/eme...
November 24, 2025 at 6:52 AM
"The very best creative people will only go to work in a few places." / "What I try to do is create the environment where these incredible people can make films." Man, this is what real leadership looks like. They don't make 'em like this anymore. youtu.be/R0XmBKsRJF8?...
Pixar's Early Days - A Never-Before-Seen Interview With Steve Jobs, 1996
YouTube video by Steve Jobs Archive
youtu.be
November 23, 2025 at 5:37 PM
Reposted by Gordon
I feel like people downplay that the major determinants of things happening in the world are:

- They are possible (eg technologically)
- They have a payout structure

Example: many forms of fraud meet 1, but not 2, and so generally don’t happen.
November 23, 2025 at 5:13 PM
Reposted by Gordon
files are actually good
the metaphor i find most useful is that your "posts", "likes", etc, are really just JSON files in this paradigm — and more concretely, concepts like "bluesky post" or "tangled repo" or "leaflet comment" are like JSON file formats. and then different apps — arbitrary even — interpret these formats
November 23, 2025 at 3:37 PM
“The tendency to favor family and friends can be overridden by rules that mandate, for example, hiring a qualified individual rather than a family member. But higher-level institutions are in some sense quite unnatural, and when they break down, humans revert to the earlier form of sociability... “
November 23, 2025 at 2:07 PM
Reposted by Gordon
once we get the culture to figure out that atproto means you can adopt new apps without selling your soul to their creators we're going to see an *explosion* of new (and old!) forms of doing shit together on the internet

it's insane how we've been held back by feudal technocratic architectures
November 23, 2025 at 5:43 AM
Reposted by Gordon
Olmo 3 is notable as a "fully open" LLM - all of the training data is published, plus complete details on how the training process was run. I tried out the 32B thinking model and the 7B instruct models, + thoughts on why transparent training data is so important simonwillison.net/2025/Nov/22/...
Olmo 3 is a fully open LLM
Olmo is the LLM series from Ai2—the Allen institute for AI. Unlike most open weight models these are notable for including the full training data, training process and checkpoints along …
simonwillison.net
November 23, 2025 at 12:17 AM
Forrester: “In the long history of evolution it has not been necessary until very recently for people to understand complex feedback systems. Evolutionary processes have not given us the mental ability to interpret properly the dynamic behavior of those complex systems in which we are now embedded.”
November 23, 2025 at 4:22 AM
Reposted by Gordon
But those are still much better problems to have! You may hate cryptocurrency but it’s really accelerated our progress in being able to secure and manage keys. Nothing solves a security problem like piles of money and constant free red teaming done by North Korea.
November 22, 2025 at 2:52 PM
Reposted by Gordon
The US is the world's largest oil and gas producer. Yet, "China is now making more money from exporting green technology than America makes from exporting fossil fuels."
China’s clean-energy revolution will reshape markets and politics
The world’s biggest manufacturer now has an interest in the world decarbonising
www.economist.com
November 7, 2025 at 11:05 AM
“adversarial poetry”
November 22, 2025 at 8:37 AM