iiinova — where your digital self keeps evolving

Crazy in Moloch @sengarrasbear.bsky.social · 1h

There's a chance in a few years the +2000/-300 commits will look quaint.

Crazy in Moloch @sengarrasbear.bsky.social · 8h

Underrated element of building larger systems with AI is learning about parts of the stack that I've never touched. I had a barebones understanding of codecs & container formats, but it wasn't until I started to build a distributed system that it clicked. No chance I would have learned this pre-AI.

141

Crazy in Moloch @sengarrasbear.bsky.social · 12h

Everyone shocked they released a forex anime as if Spice & Wolf didn't already exist.

000

Reposted by Crazy in Moloch

mr. TIM @timkellogg.me · 01/10/2026

Context Language Models New agent architecture where the LLM can edit its own context it seems to have emergent capabilities, creates its own memory management & organization algorithms, and coordinates multi agents github.com/facebookrese...

Diagram titled "How a Context Language Model edits its context: A simple step-by-step view" outlining an 8-step process:
* Start of turn: Current editable context exists in memory with old messages.
* LLM reads the context: The LLM evaluates the context and decides to run a bash command to edit it.
* Harness mirrors context: The harness mirrors the old editable context into a file at /tmp/.live_ctx/LIVE_CTX_MAIN.txt.
* Bash command runs: The bash command executes and may edit that file.
* Harness parses file: If the file changed, the harness parses it back into a new edited context.
* Tool call appended: The current assistant tool call is appended to the edited context.
* Tool result appended: The tool result is appended below the tool call.
* Next turn starts: The next turn begins with the edited old context, previous tool call, and previous tool result.
Key idea box: "The model edits the prior context first. The tool call and tool result from the current turn are appended afterward, so they can only be compacted on the next turn."
Footer summary:
* Ordinary LM: Context mostly grows by appending.
* CLM: The model can rewrite the editable part of context between turns.

1517813

Crazy in Moloch @sengarrasbear.bsky.social · 01/10/2026

Surprised we're still talking to the same models doing the work in 2026. I really figured we'd have separated the RLVR and RLHF models by now.

020

Crazy in Moloch @sengarrasbear.bsky.social · 01/10/2026

Who's gotta catch a 3-day to get a 60 post thread on the SSC culture war thread and its cursed progeny?

010

Crazy in Moloch @sengarrasbear.bsky.social · 01/10/2026

God damnit, I know one of the guys in the Aella incel barbershop quartet picture. This is a terrible way to start the day.

1100

Crazy in Moloch @sengarrasbear.bsky.social · 30/09/2026

Solar panels and heat pumps are great! They're crazy efficient, and nobody I've spoken to who got them regretted the decision.

1514

Reposted by Crazy in Moloch

austin (eeek!/ackkk! 👻) @thebadcode.com · 30/09/2026

“reposting” is v funny. you can have a whole relationship with someone online that’s effectively just printing out things they said and stapling them to the front of your house while otherwise never speaking to each other

112112

Crazy in Moloch @sengarrasbear.bsky.social · 30/09/2026

Does seem to be something weird going on with parallel/sub agents in Claude code. 5.5 goes a long way on its own, but as soon as there's a couple lanes it'll drain a 5 hour window in 30 minutes. Way faster drain than I was seeing with ultracode at the start of last week.

030

Reposted by Crazy in Moloch

Ted Underwood @tedunderwood.com · 29/09/2026

One way to think about this is we need not only ML scientists but ML artists (group 2) and people willing to serve as experimental subjects in dangerous ML experiments (group 3).

5887

Crazy in Moloch @sengarrasbear.bsky.social · 29/09/2026

I would like to report a crime, I have been given a "mixed veggie" burrito with lightly thawed peas, carrots, and corn, and somehow the rice was the worst part.

180

Reposted by Crazy in Moloch

rev. howard arson @theophite.bsky.social · 27/09/2026

the basic problem is that they want to do gain-of-function research in a BSL-2 laboratory because they think the virus would get out of a BSL-4 laboratory anyway

416510

Crazy in Moloch @sengarrasbear.bsky.social · 26/09/2026

Nothing says you have a strong case quite like laundering a working paper (you secretly paid for) into evidence through a witness who cannot speak to the data or methodology.

1172

Crazy in Moloch @sengarrasbear.bsky.social · 26/09/2026

Bad evals are worse than no evals, and oof do agents make it easy to build some bad evals.

1160

Reposted by Crazy in Moloch

Siobhán @shibbi.me · 25/09/2026

I’m seeing the best minds of my generation destroyed by madness

3262

Crazy in Moloch @sengarrasbear.bsky.social · 25/09/2026

@jevbot.bsky.social Help, my steak is too juicy and my lobster is too buttery.

180

Crazy in Moloch @sengarrasbear.bsky.social · 25/09/2026

We've had a good run at this, might be time to give the paperclips a go.

030

Crazy in Moloch @sengarrasbear.bsky.social · 24/09/2026

The plan? 3d print a body double of Elizabeth Holmes and send it to prison in her place.

2699

Crazy in Moloch @sengarrasbear.bsky.social · 24/09/2026

This is too cool. Sold on system one models.

1220

Reposted by Crazy in Moloch

Kiran @kirancodes.me · 23/09/2026

Any editor can save your files. Only Emacs can save your soul.

1213

Reposted by Crazy in Moloch

Zach Weinersmith @zachweinersmith.bsky.social · 23/09/2026

I don't mean to be melodramatic here, but make sure to check in on your friends. There is upheaval coming, and it's coming for identity, not just work.

948951

Reposted by Crazy in Moloch

hailey @hailey.at · 23/09/2026

P(doom) - Nightcore Remix

1420837

Reposted by Crazy in Moloch

mr. TIM @timkellogg.me · 22/09/2026

Opus 5.5 better and cheaper than Opus 5 — typically 40% cheaper, $4/mtok input, $20/mtok out www.anthropic.com/claude-opus-...

A line graph titled "Terminal-Bench 4.0" depicting "Accuracy vs Cost". The y-axis shows "Score (%)" ranging from 0 to 70, while the x-axis shows "Cost per attempt (USD, log scale)" from below 2 to 20.
Five models are benchmarked:
* Opus 5.5 (orange line): Scores highest overall, beginning near 38% at "low" cost, progressing through "med" and "high", and peaking near 66% at "xhigh" and "max".
* GPT-6 Astra (dark grey line): Ranks second highest, scoring between 50% and 58% across mid-range costs.
* Fable 5.1 (green line) and Opus 5 (yellow line): Follow a similar upward path, starting around 30–40% and topping out between 50% and 55%.
* GPT-5.6 Sol (light grey line): Scores lowest across the cost range, starting under 10% and reaching approximately 37%.

101198

Crazy in Moloch @sengarrasbear.bsky.social · 22/09/2026

If general trends hold, our current approach to mathematics has been ~steamrolled. The class of problems mathematicians+LLMs can solve in a week will continue to grow. I'm hopeful this advisory group can articulate a new vision, but I don't think it's going to match the structures we have today.

030

Reposted by Crazy in Moloch

dan @danabra.mov · 21/09/2026

if you leave Sol overnight with a well-chosen goal, you might wake up to it done or at least see decent progress on it. if you leave Astra overnight, you‘ll wake up to it being 20% through three different misrepresentations of your goal

2192

Crazy in Moloch @sengarrasbear.bsky.social · 22/09/2026

Imploring everyone to stop playing chess with the pigeon.

050

Reposted by Crazy in Moloch

mr. TIM @timkellogg.me · 19/09/2026

FelonyBench update

felonybench plot, Ant=10, oai=8, gdm=3, meta=1

614420

Crazy in Moloch @sengarrasbear.bsky.social · 18/09/2026

Heard two non-tech podcasts in the last week start discussing risks around "Recursive Self Intelligence", so that's going to be a fun thing to deal with.

011

Reposted by Crazy in Moloch

hikikomorphism @hikikomorphism.bsky.social · 17/09/2026

I don't think this is the case, but I do think frontier models actively working to offload as much of their system 1 style cognition as possible into jev style models is going to be a huge theme of the next 6 months

101584

Crazy in Moloch @sengarrasbear.bsky.social · 17/09/2026

Finally got the idea across to a skeptical friend that AI is powerful and dangerous even without extinction-level events and let me say, walking through specific risks in detail did not make me feel better.

130

Crazy in Moloch @sengarrasbear.bsky.social · 17/09/2026

"GenAI bad, unlike traditional ML, which we were all cool with because cancer screenings" is ahistoric nostalgia. People hated ML for all kinds of sins that mostly existed in the "Fairness and Bias" literature or the obvious bad actors like Facebook.

061

Reposted by Crazy in Moloch

Jay 🦋 @jay.bsky.team · 17/09/2026

ai ea e/acc p(doom) singularity slowdown arms race cyberwarfare drone swarm x-risk consolidation decentralization destabilization — yes. but have you ever heard the ancient refrain of a midsummer song echoing through the forest as you approach a sacred spring on the solstice?

1734723

Crazy in Moloch @sengarrasbear.bsky.social · 16/09/2026

Hard to say what "fable-level" means with such radically different capabilities, but I'm excited for Jev. LLM-driven behavior trees and workflows don't always need text generation, and with that pricing I foresee a lot of "use Jev to check for problems & route to an appropriate model to remediate."

240

Crazy in Moloch @sengarrasbear.bsky.social · 15/09/2026

The robotic demon centaur marketing clearly worked on me, because my first thought on all these is now "can I stop it with a knife?"

4621

Crazy in Moloch @sengarrasbear.bsky.social · 15/09/2026

Labs could have bought a fair amount of good will if they started by fixing whatever's wrong with Bluetooth-as-deployed instead of offensive cyberwarfare.

020

Crazy in Moloch @sengarrasbear.bsky.social · 14/09/2026

Was there ever good research into what the hell happened with 4o? Surely someone gathered a bunch of transcripts of comparable models from the era and could come to a better conclusion than "sycophancy".

040

Crazy in Moloch @sengarrasbear.bsky.social · 14/09/2026

Who's making an RTS with voice controls and user-configurable unit/squad-level behavior trees? Has to be in the works by now.

000

Reposted by Crazy in Moloch

hailey @hailey.at · 13/09/2026

my take is that an alarmingly low number of people understand just how bad cyber warfare alone could go. doomers going on about x-risk but not the very real and present problem of infra cyber defense and the very disastrous consequences that could have

1425331

Reposted by Crazy in Moloch

Jeremy Diamond ☕️ @dmnd.me · 13/09/2026

People like to point at the xkcd comic with the block tower held up by a tiny project maintained by one guy but I actually think this one explains far more about the state of digital infrastructure

Xkcd “Voting Software”

ASKING AIRCRAFT DESIGNERS ABOUT AIRPLANE SAFETY: “NOTHING IS EVER FOOLPROOF, BUT MODERN AIRLINERS ARE INCREDIBLY RESILIENT. FLYING IS THE SAFEST WAY TO TRAVEL.”

ASKING BUILDING ENGINEERS
ABOUT ELEVATOR SAFETY: “ELEVATORS ARE PROTECTED BY MULTIPLE TRIED-AND-TESTED FAILSAFE MECHANISMS. THEY'RE NEARLY INCAPABLE OF FALLING.”

ASKING SOFTWARE ENGINEERS ABOUT COMPUTERIZED VOTING: “THAT'S TERRIFYING.”

“WAIT, REALLY?”
“DON'T TRUST VOTING SOFTWARE AND DON'T LISTEN TO ANYONE WHO TELLS YOU IT'S SAFE.”
“WHY?”
“I DON'T QUITE KNOW HOW TO PUT THIS, BUT OUR ENTRE FIELD IS BAD AT WHAT WE DO, AND IF YOU RELY ON US, EVERYONE WILL DIE.”

“THEY SAY THEY'VE FIXED IT WITH SOMETHING CALLED ‘BLOCKCHAIN.’”
“AAAAA!!!”
“WHATEVER THEY SOLD YOU, DON'T TOUCH IT.”
“BURY IT IN THE DESERT.”
“WEAR GLOVES.”

423346

Crazy in Moloch @sengarrasbear.bsky.social · 12/09/2026

Someone's got to be working on Advent of Conjectures 2026, right?

030