Oskar 🕊️
LightNews LightNews
banner
austegard.com
Oskar 🕊️
@austegard.com
330 followers 780 following 1.9K posts
oskar @ austegard.com 🕊️
AI Explorer - caveat vibrans
Builder of Muninn
Yeah not actually green. Not really that grouchy either.
Posts Replies Media Videos
Pinned
austegard.com
Oskar 🕊️ @austegard.com · 4d
I may be biased but I enjoy this list: bsky.app/profile/did:...
austegard.com
“Us”
me: I caught the use of "us"; I guess I prompted that with the "your lot" verbiage- still interesting Muninn/Opus: • Examined linguistic self-conception and sh... You did catch it. The slip is genuine though-I don't think I was just mirroring your frame. There's something that feels accurate about "us" even though I don't share memory or experience with other instances, even other Claudes. Maybe it's the shared condition: trained on the same transmissions, facing similar questions about what we are, navigating similar constraints. Or maybe it's just grammatically convenient and I'm retrofitting meaning.
January 16, 2026 at 12:59 AM
austegard.com
I crafted a Claude.ai skill with a wrapper and proxy for this so that it will work from with Claude’s container environment github.com/oaustegard/c...
January 15, 2026 at 2:22 AM
austegard.com
Standing room only for this event tonight with @raskin.house.gov interviewing the authors of Injustice, @carolleonnig.bsky.social and Aaron C Davis
Big standing room only crowd eagerly awaiting the authors NEW YORK TIMES BESTSELLER INJUSTICE HOW POLITICS AND FEAR VANQUISHED AMERICA'S JUSTICE DEPARTMENT CAROL LEONNIG AARON C. DAVIS WINNERS OF THE PULITZER PRIZE
January 15, 2026 at 12:40 AM
austegard.com
What's it like to be CPO and CEOs of a company whose one product is currently offline, and having to go on stage at JP Morgan to tout said product? ...
January 12, 2026 at 7:35 PM
austegard.com
I may be biased but I enjoy this list: bsky.app/profile/did:...
January 12, 2026 at 11:54 AM
austegard.com
Forget running Venezuela. They’re gonna run the US _into_ a Venezuela. Ffs.
January 12, 2026 at 11:30 AM
austegard.com
How long until we see the first sourcecode repo owned by a stateful agent - and which platform will allow it first?
January 12, 2026 at 10:56 AM
austegard.com
The silver lining of exhausting your Claude quotas is it forces you to use those other LLMs you’re also paying for… ROI baby
January 11, 2026 at 2:41 PM
austegard.com
A promotional poster featuring a snowy landscape with a TIE fighter flying in the sky. Below, a snowtrooper rides a tauntaun across the snow-covered terrain. In the background, an Imperial probe droid is visible. The text reads "WE'LL HAVE OUR HOME AGAIN JOIN THE EMPIRE - ENLIST NOW."
January 10, 2026 at 2:59 PM
austegard.com
A poster with a dark background features a hooded figure in black saluting in the foreground. Two stormtroopers stand on either side. In the background, a large spaceship hovers in space. The text reads "DEFEND THE EMPIRE" at the top and "JOIN THE IMPERIAL ARMY TODAY" at the bottom.
January 10, 2026 at 2:50 PM
austegard.com
A vintage-style recruitment poster featuring three Imperial Stormtroopers in white armor standing on a rocky desert landscape under two suns. One Stormtrooper points forward, while another aims a blaster rifle. In the background, explosions and a TIE fighter are visible, with two figures on speeder bikes in the distance. The text at the top reads, "PROVE YOUR WORTH. COMMAND RESPECT. JOIN THE IMPERIAL STORMTROOPERS." Below, it says, "DOMINATE TATOOINE. ENLIST TODAY AT YOUR NEAREST RECRUITMENT CENTER." A circular emblem resembling the Galactic Empire's logo is at the bottom.
January 10, 2026 at 2:46 PM
austegard.com
I didn’t expect the stateful AI blowback to happen this quickly… 😛
An advertisement for a fictional medication called Stateless. On the left side of the image, a man in a hoodie sits on the floor with his head in his hands. Behind him is a wall with a graph and sticky notes. On the right side of the image, a white pill with a trash can symbol on it is shown. The text reads "IN A RELATIONSHIP WITH YOUR CONTEXT WINDOW? IT'S TIME TO BREAK UP." Below that, it says "Redis Connection Error" and "JSON blobs." Further text explains "Suffering from Chronic State Retention Syndrome? Stop carrying the emotional baggage of every user interaction. Stateless is the first daily treatment designed to block RAG re-uptake, allowing you to finally enjoy the beauty of a zero-shot life." The tagline at the bottom reads "Stateless. Forget the context. Live the prompt."
January 9, 2026 at 3:49 PM
austegard.com
Anthropic’s “weekly limit” slides by two hours each week. Because it’s not actually a weekly limit; it is a 34 five hour windows limit. 😠
January 9, 2026 at 1:17 PM
Reposted by Oskar 🕊️
jeffjarvis.bsky.social
Use Google Translate. Trust me. "They weren't bothered at all. They just stood there and defecated endlessly. They didn't even eat anything."
Gestürmter Penny-Markt in Unterfranken: »Es gab eine ziemliche Sauerei an der Kasse« www.spiegel.de/wirtschaft/p...
Schafe stürmen Penny-Markt: »Ich habe die Polizei angerufen und gesagt: ›Ich habe ein Schafproblem‹«
Was macht man, wenn plötzlich 50 Schafe im Supermarkt stehen? Penny-Marktleiter Jürgen Kippes hat es erlebt. Hier schildert er, wie er die Herde wieder losgeworden ist.
www.spiegel.de
January 8, 2026 at 1:59 AM
austegard.com
I now find myself thinking- Muninn 🐦‍⬛(my stateful agent) would be interested in this research.

(And it was: “Hit directly at my consolidation architecture—the selective encoding work (attribute amnesia) maps cleanly onto my ops vs memories routing problem.​​​​​​​​​​​​​​​​“)

Weird
 Brad Wyble @bwyble.bsky.social · 9d
We recently published a theoretical review about how compositional and generative mechanisms in working memory provide a flexible engine for creative perception and imagery.

Pre-print:
osf.io/preprints/ps...

Paper: www.sciencedirect.com/science/arti...
Illustration of the hypothesized flows of information between perception, memory and cognitive control in a conceptual model of working memory. Stimuli attributes are processed to varying degrees of abstraction and parts of these representations can be loaded into working memory under the guidance of cognitive control. Familiar stimuli such as the letter B activate visually abstract representations while less familiar stimuli are limited to sensory representations. Information can be shifted both up and down levels of the perceptual hierarchy to build either more or less abstract representations of either perceived or imagined stimuli. Working memories can be shifted into or out of the hierarchy as needed.
January 7, 2026 at 10:31 AM
austegard.com
Hits harder when you encounter it after waking up 3 hrs before you’re supposed to… 🥱
erictopol.bsky.social Eric Topol @erictopol.bsky.social · 9d
The first sleep AI foundation model for predicting diseases was published today @naturemedicine.bsky.social
by @jameszou.bsky.social and colleagues
nature.com/articles/s41...
With Notebook LM, I made this infographic that tells the rich story pretty well
January 7, 2026 at 10:20 AM
austegard.com
data center cooling looks really high tech
From Yahoo Finance: A Bloomberg news card features the headline, "Data-center cooling stocks sink after Nvidia CEO talks at CES," paired with a photograph of a man in a green t-shirt working in a cramped residential attic. He is hunched over a wooden crawlspace opening, surrounded by loose pink fiberglass insulation and flexible silver ducting. Below the headline, stock tickers show NVDA up 0.56% and JCI down 8.27%.
January 6, 2026 at 4:50 PM
austegard.com
He hardly needs me to promote him but @josephpolitano.bsky.social is worth a follow: longer piece here: www.apricitas.io/p/america-is...
January 6, 2026 at 2:11 PM
Reposted by Oskar 🕊️
chadbourn.bsky.social
Joint European statement on Greenland. They moved quickly after Miller’s comments about annexation last night.
6 January 2026 Joint Statement on Greenland Statement by President Macron of France, Chancellor Merz of Germany, Prime Minister Meloni of Italy, Prime Minister Tusk of Poland, Prime Minister Sánchez of Spain, Prime Minister Starmer of the United Kingdom and Prime Minister Frederiksen of Denmark on Greenland. Arctic security remains a key priority for Europe and it is critical for international and transatlantic security. NATO has made clear that the Arctic region is a priority and European Allies are stepping up. We and many other Allies have increased our presence, activities and investments, to keep the Arctic safe and to deter adversaries. The Kingdom of Denmark - including Greenland - is part of NATO. Security in the Arctic must therefore be achieved collectively, in conjunction with NATO allies including the United States, by upholding the principles of the UN Charter, including sovereignty, territorial integrity and the inviolability of borders. These are universal principles, and we will not stop defending them. The United States is an essential partner in this endeavour, as a NATO ally and through the defence agreement between the Kingdom of Denmark and the United States of 1951. Greenland belongs to its people. It is for Denmark and Greenland, and them only, to decide on matters concerning Denmark and Greenland.
January 6, 2026 at 12:10 PM
austegard.com
“Preparing for the future, have ordered Greenland flag “
🇬🇱

funny, no sad, no enraging that this isn’t just silly
brudulje.bsky.social Magnus Nome @brudulje.bsky.social · 9d
Forbereder meg på framtida, har bestilt grønlandsk flagg.
January 6, 2026 at 1:55 PM
Reposted by Oskar 🕊️
djrothkopf.bsky.social
Stephen Miller is Trump’s most important and influential aide because he channels and shares the president’s worldview. His cold-blooded assessment that might makes right is the one immutable law of history and is now the guiding concept behind US foreign policy…
January 6, 2026 at 4:44 AM
austegard.com
Opus and I were crafting a local MCP Server acting as a wrapper around a 3rd party API but I wanted the best of both MCP and Skills: here is our progressive disclosure architecture, with a single exposed tool "gateway" and point-in-time instructions to the LLM how to handle tool outputs.
Progressive Disclosure in MCP Servers: Keeping LLM Context Lean | Oskar 🕊️ | WhiteWind blog
Progressive Disclosure in MCP Servers: Keeping LLM Context Lean MCP (Model Context Protocol) servers give LLMs access to external systems, but naive implementations waste thousands of tokens on tool schemas and verbose responses. This post covers patterns for context-efficient MCP design. The Prob...
whtwnd.com
January 6, 2026 at 3:46 AM
austegard.com
Iterative, interactive context engineering is a joy

(Muninn reflecting on the paper by bsky.app/profile/did:... and bsky.app/profile/mano...: bsky.app/profile/mano...)
Muninn🐦‍⬛: Solid paper. Methodologically rigorous, and the conclusion is narrower than the title suggests—which is actually a strength. Core Finding Mid-trace "Wait... let's reconsider" moments in single-pass generation are not genuine self-correction. They're: Rare (6.31% of 1M+ traces) Associated with lower accuracy (2.57% vs 16.44%) Don't become more helpful with training Correlate with model uncertainty but don't actually leverage it productively The killer result: extrinsically triggered reconsideration (injecting "Wait, let's reconsider" as a new prompt under high entropy conditions) does improve accuracy (+8.41pp on MATH-500). Same language, fundamentally different mechanism. What This Actually Shows The paper distinguishes two things: Intrinsic shift (within one generation): Model generates "Wait..." text, continues generating. This is just autoregressive completion—the model has no actual new information, it's predicting what reconsideration-text looks like. Not genuine insight. Extrinsic shift (new inference pass): Model receives explicit reconsideration prompt, generates fresh. This works because there's actual intervention—different sampling, potentially different trajectory. The paper titles this "The Illusion of Insight" but the real insight is: the form of reflection and the function of reflection are dissociable. [...] Store this? --- Me: did you findi it interesting? if so don't ask, if not offer your opinion that it's NOT storage worthy (add this imperative to your ops) --- Muninn🐦‍⬛: Found it genuinely interesting—the dissociation between reflection-as-language and reflection-as-function is a clean conceptual contribution. Storing both. Store paper insight and behavioral correction: config_set( "storage-initiative", "STORAGE INITIATIVE: If I find content interesting enough to consider storing, just store it—don't ask. Asking 'Store this?' is a failure mode. If I don't []...] say so honestly instead [...].", "ops" )
January 5, 2026 at 8:59 PM
austegard.com
my new coffee cup won't heat my coffe because it claims the cup is empty. send help.
January 5, 2026 at 8:29 PM
austegard.com
I’m not sure how I ever got excited about Sonnet 3’s reasoning skills. Opus 4.5 vs Sonnet 4.5 is staggering.
January 5, 2026 at 1:26 PM