Sometimes really don’t get why a company like Partiful can’t fix the bug of don’t know which city is the event before registration? It’s been years now
I’ve been researching depth-residual routing & delta memory- Kimi K3 proved this holds at 2.8T params. But only for models trained this way
I then asked: can you retrofit this onto existing LLM/world models?
Introducing Retro-DARC: a Delta Attention Residual Compute adapter
that retrofits depth-selective memory onto existing models without retraining:
EA exact no-op at insertion (bitwise 0.0 deviation on 3 public checkpoints - proven exact even through MoE routing and early exit), adapter-only training, exact rollback, and content-level memory audits that weight-touching adapters can't express by construction.
The early results are promising. A ~262K-param adapter beat LoRA on a frozen public checkpoint, inserted at zero behavioral risk. And it's the first memory you can audit: zero the bank and the frozen loss returns to machine precision; shuffle it and 70–107% of the gain vanishes. That's a model measurably using its own computation history.
Retro-DARC is useful for improving K3 too. K3 serves in MXFP4/MXFP8, and layer outputs - exactly what its AttnRes retrieves - collapse under that corruption in my tests (0.04–0.75 top-1) while normalized delta keys hold 0.86–0.99. A Softmax1 null route with proven mass bounds lets the router decline to retrieve instead of forcing the sinks and outliers OASIS documented; typed memory turns "what is depth retrieval doing?" into a runnable audit. And one experiment is free: K3 already computes a per-step KDA update magnitude and throws it away - that's a depth-saliency prior waiting to be read.
For world models, the same contract becomes one typed memory interface: Retro-DARC-X gives world/action planners a single bank - observations, actions, tools, latent causes, traces, physics residuals, failures - the substrates a planner needs to remember in order to act over long horizons in the physical world.
You can experience the thesis yourself. I created https://t.co/XCE4IN3jHK which runs the paper's memory contract inside a walkable AI world. Innovation Memory makes your interactions survive looking away or reloading, and Dream Residue turns your movement into an inspectable glowing trace.
Join the main https://t.co/CwGJgV8H91 co-creation exhibition - the contest opens Wednesday. One of the six memory applications of the paper - saliency-weighted curation, will rank worlds by genuine novelty rather than recency and pick each world's most surprising viewpoint. And we're bringing live shows to SF + LA next month: screens showcasing worlds that remember, so you can literally play hide-and-seek with a dream that knows where you hide. Hide-and-seek is the exact test every 2026 world-model benchmark says the field fails, something leaves the frame, keeps changing, and has to come back right.
This has been a fun little one-person research project so far! I’d love feedback and keep working on it. I’ve had over 100 references- thanks for all the prior work from teams that inspired me!
I'd also love to intern at a GPU-rich lab (lol) and keep making beautiful worlds - if you let me keep running my remaining tests :)
2MB adapter kit, demo, prior iterations:https://t.co/n8VXPZDtYr
Paper: https://t.co/avxjL1Ekwj
Economist Zhu Jiaming, on the Faustian bargain of the AI age:
“What we should truly fear is that humans surrender decisions because thinking is exhausting, surrender love because emotion is troublesome, and surrender governance to algorithms because it’s too complex.”
The real tragedy of AI may not be that machines replace us, but that we hand over the parts that make us human.
if you're a performance marketer, here's how I use a custom Claude Cowork plugin to manage Google Ads at @AnthropicAI. it connects to the Google Ads API via MCP, encodes my common paid search workflows into skills, and works on desktop and Dispatch.
Launching https://t.co/23kQj6OkUu, a database for founders applying to accelerator programs.
> Deals tab: compare accelerator terms in 30s
(A16Z/YC/Techstars/Antler/South Park Commons...)
> Credits tab: find free API/cloud credits (Cloudflare/AWS/Google Cloud/Vercel)
On average, accelerator deals imply your new startup is worth about $2.2M.
So everyone needs to get the private key level security protocol with family now. Literally the whole crypto military-level security is just a pre-training for the human battle with AI!!!!
My wife calls me, panicked.
The call is from her number, and her voice is unmistakable- that’s my wife.
‘Babe, our son is hurt. He got in a bike wreck. I’m at the emergency room but they won’t take our insurance and I need cash to get him help. Please send me 3000 dollars as soon as you can, he’s really not doing well.’
Me- ‘Wow, that’s scary. Tell me our passphrase and then I’ll send the money.’
Her (it) - ‘What? What passphrase? This is your wife, our son is hurt. Send the money now!!’
Me- ‘I’ll call you back. I don’t believe that this is my wife. If it is, I’m sorry, but we discussed this.’
The number? Spoofed. Easy to do and there’s no way to tell if a phone number is being spoofed aside from hanging up and calling back to confirm.
The voice? AI generated. Easily done. A few seconds of audio is all it takes to create a realistic audio deepfake.
What can you do?
1) Create a family safe word or passphrase. Ours is definitely not ‘Keep Going’ although we considered it. Discuss the passphrase far away from phones or any recording device. This is as analog as possible. Don’t forget that the trigger for the passphrase is just as important as the phrase itself. So instead of asking ‘what’s the safe word?’ have a separate triggering question. For example, you could say ‘I’m eating banana cream pie’ and this would trigger your spouse to respond ‘purple velvet pillows’ if that’s the safe word.
Make it fun, silly, and easy to remember. And DON’T WRITE IT DOWN.
2) Cognitive security is an essential skill in 2026. Assume every image and video you see online is fake until proven otherwise. Expect scams and spammers, and be pleasantly surprised when it’s not.
3) Figure out a backup communication option with people who you absolutely need to be able to reach. Don’t just rely on a phone number for communication. Have redundant, ideally encrypted methods of communication with family.
What did I miss? I think (hope) Nikita is wrong on the timeframe- agentic bots like Claude bot are impressive but not quite ready to flood the phone lines in just 90 days. But I think it’s going to be a huge problem by the end of the year. I already get dozens of increasingly realistic spam calls and texts daily- it’s only going to get more annoying. Have a plan to keep your family and your finances safe!
Curious and excited to see what kids will be passionate about coding in the near future. They can create games soon, and we should play.
They have the purest imaginations. it’s the ultimate unconstrained sandbox. Our brains are essentially optimized prediction machines. For an adult, a hallucination is a bug; for a child, it’s a feature. We should optimize for The Beginner's Mind.