Introducing RePo: Language Models with Context Re-Positioning Standard LLMs force a rigid linear structure on context, treating physical proximity as relevance. Cognitive Load Theory suggests this is inefficient—models waste capacity managing noise instead of reasoning. arxiv.org/abs/2512.14391

NBA game in London. Someone screams “Leave Greenland alone!” during US national anthem. The crowd reacts in support.

Wow. CC limit on the Pro plan feels very short now… (decided to call a temporary bankruptcy for apchat and get some help, as it was just making circles). And I didn’t even use Opus. Might give a try to GLM-4.7 from z on the same “bankruptcy point” and if it performs well, cancel altogether.

This is sick

“Please make the tests pass” gets me this. My fault… they technically do pass now!

Not only humans are suffering From LLM hallucinations… :) (the llm_oneshot is a pure llm call, with no access to files or tools).

Ooh. Ollama just released Anthropic API impersonation, meaning you can set the endpoint to a local port and be off to the races! docs.ollama.com/integrations...

An idea for the new feature: time-based LLM API change. Because I don’t wanna sleep next to a working GPU… and the second rig thats away is much less powerful… this can be decomposed into time-based event injection *and* a tool to allow LLM set its own API URL. But, I bet this would distract.

🙋‍♂️Hello to all of my new followers! :) My AI tinkering at the moment is mostly with my own half-baked agent that I called APchat, trying to figure out whether I can get meaningful results off the smaller local models. Currently - Devstral-2-Small. It’s a bumpy journey, so buckle up! :-)

Interesting that despite the near-identical top-level “metaprompts”, the commit discipline from run to run is very different - sometimes it is really discrete commit progression, and sometimes there is no commits at all.

Some pics. The numbers on the left side are PIDs of child processes. Let’s see if with 4 levels of nesting it might one-shot it with no interaction required..

The summary for the day: - curly file glance tool implemented. the result looks a bit buggy but “good enough”. - integrated the first cut of tools to deal with persistent memories (I used a different project which I temporarily placed in ./tmp as a “source of inspiration” for my APchat to look at.

Came up with a new idea for a tool: “curly file glance” tool: give a file name and optional start line, and the tool would read the lines, elide everything that is between the top level curly brackets, and print the stuff preceding the opening curly brackets.

More than 200,000 Danish citizens have signed a petition to buy California as a response to Trump’s attempt to take Greenland. They say they will provide Californians with “rule of law, universal health care, fact-based politics, and a lifetime supply of Danish pastries.”

Is this the moment we can finally get serious about furthering democracy, structuring markets so they work for people, and getting started on tech sovereignty? We know how and we have the means. Trade appeasement hasn't worked. Let's get moving! 🇪🇺

interesting - I did not realise half of this bylinesnetwork.substack.com/p/decline-wh...

> Europeans enjoy more free time than Americans, a higher life expectancy and lower levels of inequality – all with roughly comparable productivity.

We've learned that for the All Lives Matter crowd, no lives matter.

grumble. Am I the only one using Bsky on Safari on the iPhone ?

The risk of AI for education is not students cheating in exams, it is people in general cheating themselves into believing they understand things they don’t.

Why are In-n-Outs draining the oceans????

Ukraine now has not a single power plant that was not hit by missiles, most of them not operational right now. Kiev currently is -13 °C. People are freezing for days, pipes are bursting. And we got so used to the news of the war, that it's not even a front page story any more.

I'm surprised the extent to which mature, secure protocols like XMPP have not yet been adopted by the agent community Most of the venues I've sent folks recommend have nice API's, but are all 3rd party owned or not private Whereas e.g. XMPP is proven, free, and easy to use

github.com/ayourtch-llm... - 9 automagic commits later, the feature works. Again, I would fail this 22b local model on architectural grounds if it were Opus. But… it’s local. More forgivable.

if I were to judge by the paperwork/code ratio - this is an enterprise grade solution ! 😂

In today’s uncertain and worrying world, it’s always uplifting to read some positivity and 100% confidence 😂😂😂

LLM empire building. even if I asked it to do 2 levels, it went all the way with 4 levels of delegation - the last two being seemingly identical. If this repeats I will probably have to add a performance review or something 😂

The only environmental impact of my yesterdays tinkering with LLMs was that I had to turn down the heating: it was all local. Let’s see if I can keep it up today “off the grid” as well. The ”director agent” approach does appear to slice complexity rather well, to a level of not needing compaction.

So, the “director agent“ pattern seems to be pretty effective. Got it to implement (with a few hiccups) the tool output length limiter (so it doesn’t eat the content window) - from the first shot as seen from “director agent”. I didn’t see at all what the manager and worker levels were doing.

I am calling for a complete and total boycott of the Mercator projection in all news stories about Greenland until every member of the American public has seen this

I wrote a complaint to Apple about the Apple App Store hosting X, the non-consensual porn and child-abuse image generator and distribution app. This was the reply I received:

I ask my coding agent to launch a manager subagent which would plan the work based off an existing high level plan and oversee its implementation by sub-sub-agents - under an assumption that the “manager agent” will have the knowledge of the plan but not the minute details of the code.

If a subagent is launching another subabent, does the original agent get promoted to a director agent now or only senior manager ? 🤪

Playing with huggingface.co/unsloth/Devs... and it feels quite “crisp” so far. Especially after the exhaustingly chatty Nemotron-30B. It made several edits to apchat which worked immediately, and for some others managed to get itself out of the woods with Rust compile errors.

🇧🇪 Journalismfund Europe announces the launch of Belgian Bridge. This new grant programme aims to strengthen connections across Belgium's linguistic divide through collaborative investigative journalism. 🗓️ Deadlines: 26 February & 25 June 2026 More info: www.journalismfund.eu/news/belgian...  #BEB

One more thing to try when I get back to the keyboard…

Spotted this article on HN and wondered: what if one puts these tokens somewhere in the beginning of the files that coding agents deal with ? Would they affect the processing of the document by LLM? Or by virtue of being “close to the center” they would be very effective No-ops ?

the one weird trick to counter the massive propaganda apparatus arrayed against Zohran: just keep doing good shit every day

y'all say writing code is hard, try ansible playbook/task spaghetti inside multiple interdependent GitHub Actions... :)

Boy is Nemotron-30B-A3B model wordy. Seems to more or less get things done eventually, but.... so. much. talking. :D

Inspired by @dollspace.gay’s VDD methodology, I’ve created a Letta agent called Crucible that will critically examine your code to discover flaws: app.letta.com/agentfiles/a.... As a memory agent, it also has the opportunity to learn new flaws and patterns, and extend its criticism to other domains.

Excellent

My little guinea pig coding agent has decided that instead of implementing a feature it’s better to delegate the whole thing to a subagent. 🤷‍♂️ Looking forward to see the results…

Some people are arguing AI agents are bad. Based on what I am seeing human agents are much worse.

spotted on LinkedIn : github.com/tuned-org-uk... - too long of a note to put here, so I will not try :-)

The Epstein files. Whatever happened to them?

When Renee Good said "I'm not mad at you" what the ICE goon heard was "I'm not scared of you." That's what enraged him. The populace wasn't fully terrorized and cowed into submission yet, so he acted to make that happen.

Sometimes you just have to let @lopatto.bsky.social say it out loud www.theverge.com/policy/85990...

BIG FIGHT AT POLAND’S WROCLAW ZOO!!!

The thing about fragile masculinity and all the things marketed to those susceptible to it is that there far cheaper ways to be all alone at the end of the evening, masturbating into a sock, so why not just save your money