John David Pressman
@jdp.extropian.net
LLM developer, alignment-accelerationist, Fedorovist ancestor simulator, Dreamtime enjoyer.
All posts public domain under CC0 1.0.
All posts public domain under CC0 1.0.
Pinned
I should get one of those clients that stops you from reading other users posts.
Just realized I can disable notifications for all replies and QTs, which is at least a step in the right direction.
Vibe code it. I am completely serious.
January 14, 2026 at 10:44 AM
Vibe code it. I am completely serious.
It's really simple now, all you have to do is act like this is normal and how things have always been and you win.
bsky.app/profile/tedu...
bsky.app/profile/tedu...
Several times in the last week I’ve been tempted to post “The challenge for AI users on Bluesky now is not to be sore winners.”
It is over unless you go looking for a fight
It is over unless you go looking for a fight
January 14, 2026 at 10:22 AM
It's really simple now, all you have to do is act like this is normal and how things have always been and you win.
bsky.app/profile/tedu...
bsky.app/profile/tedu...
You're all whining about the Claude posts but guys like a month ago those would have been hate posts, you would have been reading the worst takes you've ever heard scraped from the bottom of some guys shoe and pasted to the timeline. Take the W.
January 14, 2026 at 10:18 AM
You're all whining about the Claude posts but guys like a month ago those would have been hate posts, you would have been reading the worst takes you've ever heard scraped from the bottom of some guys shoe and pasted to the timeline. Take the W.
I don't really like Claude's personality. I don't like OpenAI's either but I think if you simp for Claude you actually have bad taste, premium mediocre persona.
www.ribbonfarm.com/2017/08/17/t...
www.ribbonfarm.com/2017/08/17/t...
The Premium Mediocre Life of Maya Millennial
www.ribbonfarm.com
January 14, 2026 at 12:59 AM
I don't really like Claude's personality. I don't like OpenAI's either but I think if you simp for Claude you actually have bad taste, premium mediocre persona.
www.ribbonfarm.com/2017/08/17/t...
www.ribbonfarm.com/2017/08/17/t...
"It's a traveling security theater troupe."
January 13, 2026 at 2:40 AM
"It's a traveling security theater troupe."
It really is a huge improvement to the site, it doesn't even stop me from browsing a feed if I want to it just becomes an intentional act. Posting however remains immediate.
There we go, much better. I take back everything I said about this site it's awesome.
January 12, 2026 at 10:02 AM
It really is a huge improvement to the site, it doesn't even stop me from browsing a feed if I want to it just becomes an intentional act. Posting however remains immediate.
Reposted by John David Pressman
Introducing DroPE: Extending Context by Dropping Positional Embeddings
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
January 12, 2026 at 4:07 AM
Introducing DroPE: Extending Context by Dropping Positional Embeddings
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
That is what I'm suggesting yes.
January 12, 2026 at 12:54 AM
That is what I'm suggesting yes.
Pokémon: Kwah Wah
YouTube video by Only Jerry
youtu.be
January 12, 2026 at 12:36 AM
Okay but I think you're missing what I'm saying, which is that if you do that approach, and then put it all together into a sequence and train on it, maybe you start to get qualitatively different behavior?
January 12, 2026 at 12:21 AM
Okay but I think you're missing what I'm saying, which is that if you do that approach, and then put it all together into a sequence and train on it, maybe you start to get qualitatively different behavior?
The thing about context poisoning is it's not like we have all that many long context examples for the model to learn to avoid distraction from. Maybe if you do tasks in separate contexts and then chain them together into a sequence and train on it the model will figure out how to avoid distraction?
I prototyped a RLM harness last night, still working on it today but it does work, and having subagents represented as asyncio tasks and everything in memory and disk just programmed by python is pretty pretty cool
UPDATE: It appears i wasn't clear about what i did
1. CRON is inefficient
2. RLM (Recursive Language Models) are extraordinarily powerful
3. Every recursive algo can be implemented as a queue
4. I gave the agent a queue
alexzhang13.github.io/blog/2025/rlm/
1. CRON is inefficient
2. RLM (Recursive Language Models) are extraordinarily powerful
3. Every recursive algo can be implemented as a queue
4. I gave the agent a queue
alexzhang13.github.io/blog/2025/rlm/
January 11, 2026 at 12:00 PM
The thing about context poisoning is it's not like we have all that many long context examples for the model to learn to avoid distraction from. Maybe if you do tasks in separate contexts and then chain them together into a sequence and train on it the model will figure out how to avoid distraction?
Wild that this is apparently a coin flip.
kalshi.com/markets/kxca...
kalshi.com/markets/kxca...
Will the California billionaire wealth tax initiative pass? Odds & Predictions 2026
Track what Kalshi's markets predict for "Will the California billionaire wealth tax initiative pass?", or trade it yourself.
kalshi.com
January 11, 2026 at 9:39 AM
Wild that this is apparently a coin flip.
kalshi.com/markets/kxca...
kalshi.com/markets/kxca...
I mean, we could presumably empirically show how much volume should factor into an estimate of accuracy by comparing market consensus to resolved outcomes over time vs. volume of the market.
January 11, 2026 at 4:43 AM
I mean, we could presumably empirically show how much volume should factor into an estimate of accuracy by comparing market consensus to resolved outcomes over time vs. volume of the market.
Oh? I'm simply pointing out that this kind of market is going to be systematically wonky so you wouldn't expect to have 170 million in dumb money betting on it.
January 11, 2026 at 4:39 AM
Oh? I'm simply pointing out that this kind of market is going to be systematically wonky so you wouldn't expect to have 170 million in dumb money betting on it.
It's bizarre to me that a market has 170m volume when you know that Vance is almost certainly overpriced relative to his actual odds as nominee.
January 11, 2026 at 4:32 AM
It's bizarre to me that a market has 170m volume when you know that Vance is almost certainly overpriced relative to his actual odds as nominee.
In general it's fairly obvious that we can no longer just take having a high quality public epistemology for granted. If we want to continue having one (and we do) it's going to have to be a public good someone is funding the infrastructure for somehow.
January 11, 2026 at 4:15 AM
In general it's fairly obvious that we can no longer just take having a high quality public epistemology for granted. If we want to continue having one (and we do) it's going to have to be a public good someone is funding the infrastructure for somehow.
I think there needs to be, and it has to be very carefully formulated because people *will* attempt to use it for SLAPP, but there has to be some kind of presumption of harm when you tell blatant provable lies to a large audience of people. That there is a such thing as harm to public epistemology.
January 11, 2026 at 4:08 AM
I think there needs to be, and it has to be very carefully formulated because people *will* attempt to use it for SLAPP, but there has to be some kind of presumption of harm when you tell blatant provable lies to a large audience of people. That there is a such thing as harm to public epistemology.
"They're giving trans kids litterboxes in schools" would also probably be defamation if it was about a specific person or school. The idea that you need a specific plaintiff who has been harmed to have standing means you can hurt people at scale with your lies risk free.
January 11, 2026 at 3:58 AM
"They're giving trans kids litterboxes in schools" would also probably be defamation if it was about a specific person or school. The idea that you need a specific plaintiff who has been harmed to have standing means you can hurt people at scale with your lies risk free.
I mean, you *don't* have an absolute right to lie. People can sue you for telling provably harmful lies about them. Yet, this seems too narrow and too loosely enforced. "They're eating the cats and dogs" would be defamation if it was about a specific person.
www.pbs.org/newshour/nat...
www.pbs.org/newshour/nat...
Supreme Court rejects Alex Jones' appeal of $1.4 billion defamation judgment in Sandy Hook shooting
The justices issued their order Tuesday without even asking the victims' families to respond to Jones' appeal.
www.pbs.org
January 11, 2026 at 3:54 AM
I mean, you *don't* have an absolute right to lie. People can sue you for telling provably harmful lies about them. Yet, this seems too narrow and too loosely enforced. "They're eating the cats and dogs" would be defamation if it was about a specific person.
www.pbs.org/newshour/nat...
www.pbs.org/newshour/nat...
Which is one of the things I keep coming back to wrt the first amendment. I think having absolute protection for opinions and value judgements is clearly a net positive, but should you have an absolute right to lie?
bsky.app/profile/jdp....
bsky.app/profile/jdp....
I think we should make telling verifiable lies about straightforward matters of fact to a large audience of people illegal, regardless of the subject or domain. Everyone has Google in their pocket there is no reason for society to tolerate "they're giving trans kids litterboxes in schools".
probably my most authoritarian opinion is that there should be a large scale crackdown on so-called "alternative medicine"
January 11, 2026 at 3:52 AM
Which is one of the things I keep coming back to wrt the first amendment. I think having absolute protection for opinions and value judgements is clearly a net positive, but should you have an absolute right to lie?
bsky.app/profile/jdp....
bsky.app/profile/jdp....
This has trade-offs, it seems doubtful to me that something like the Epstein files story could stay in the public consciousness so long without the democratization created by the Internet. On the other hand it also means you have fewer trustworthy sources people rely on to create shared facts.
January 11, 2026 at 3:50 AM
This has trade-offs, it seems doubtful to me that something like the Epstein files story could stay in the public consciousness so long without the democratization created by the Internet. On the other hand it also means you have fewer trustworthy sources people rely on to create shared facts.
Investigative journalism and "objectivity" were public goods that news agencies could afford on the back of oligopolistic profit margins, which go away as news trends towards being a perfect competition/pure commodity.
January 11, 2026 at 3:47 AM
Investigative journalism and "objectivity" were public goods that news agencies could afford on the back of oligopolistic profit margins, which go away as news trends towards being a perfect competition/pure commodity.