Satwik Panigrahi

Wrote this for @CalBlockchain when I first joined. Super cool protocol called @3janexyz

OpenAI scaling its Trusted Access for Cyber with GPT-5.4-Cyber is obviously a response to Project Glasswing. However, what I find more interesting is the irony of it all. OpenAI now champions (kinda) democratized to specialized tools. Meanwhile Anthropic, who has always had an open ethos, opts for a more gated frontier capabilities through closed programs.

Continuous learning, autoresearch, KV-cache compression, etc. All this is ways to help agents master forgetting, abstracting, and rebuilding latent awareness. That's where generality emerges.

What results from this is a self-perpetuating RL system that continuously learns its internal model, adapts to new environments with minimal overhead, and scales intelligence without shear token usage.

All the latest architectural "unhobblings" on the agent side keep pointing to the same truth: context bloat is the enemy. And while most commentary points to the benchmarks, I see something more fundamental at play.

@limitlesstack @isaakfreeman short version is that there are humans with DEC2/ADRB1/NPSR1 variants who naturally sleep ~5 hrs and thrive. the real target is the orexin pathway with the first direct orexin agonists (TAK-861) on trial. this prolly will be the way we get there, not oral peptides lol.

@limitlesstack the question tim urban is actually asking of engineering sleep compression is a legitimate and fascinating one. if you want to see what serious thinking on this looks like, @isaakfreeman wrote this isaak.net/sleepless

a pill that cuts your need for sleep from 8 hours to 6 while making those 6 feel like 9 would be the most valuable compound on earth it exists. the pinealon + epitalon combo. pinealon: - improves sleep architecture quality - longer deep sleep, more REM - less waking mid-cycle - zero brain fog in the morning - reduces neuronal oxidative stress - effects build over 20-30 days and persist weeks after you stop epitalon: - anti-aging via telomere lengthening at the cellular level - directly stimulates melatonin synthesis in the pineal gland - regulates circadian rhythm - restores function after pineal gland calcification - reduces cortisol levels - roughly 40% reduction in mortality among elderly in a 6 year study and they ACTUALLY work in oral form. no injections needed. as ultrashort peptides they are small and stable enough that they pass through the gut intact. dose + cycling: 2 pinealon capsules. 1 epitalon capsule. everyday, 30min before bed. run it 20-30 days straight. then stop for 3 months. repeat. optimal cycle timing is very individual. use this as a baseline, then trust your gut and adjust accordingly. only vendor i use: us domestic epitalon yourprotocol.co/products/epita… us domestic pinealon yourprotocol.co/products/pinea… not medical advice.

The best skills is to learn how to quickly figure out what should be built, why, and how to integrate it into the real world.

I just listened to @gokulr talk about this — PMs are now hands-on prototypers that decide the evals for non-deterministic code. The new moat is therefore judgement + outcomes + integration.

The software engineering landscape is no longer "how fast can we code this?" The new constraints will be: 1. How fast can we decide WHAT to build? (vision, strategy, priorities) 2. How do we integrate these systems into messy legacy stacks without everything breaking?

I assume a good future direction is looking to applying the similar normalization heuristic to JEPA. It could enhance latent predictions and reduce energy mismatches.

Similarly GDPO decoupled reward normalization which led to sharper preference optimization and stable training. Nothing flashy, yet leads to compounding returns. With more stable training, a lot of these vision tasks will lead to stronger transfer to world simulaation.

A pattern I've been seeing lately is that most research boil down to clever normalization tweaks, which from first principles, makes sense. Recently ViT-5 achieve massive benchmark advancements off of replacing LayerNorm for RMSNorm + LayerScale.

@gregisenberg With a growing agent economy, I think this is where things like x402 and a lot of the blockchain space will go to. @NEARProtocol has a lot of cool ideas of how TEEs and x402 can be used to prove the credibility of an agent.

startup idea for you - linkedin for ai agents linkedin sold for $26.2b in 2016, what is the linkedin for ai agents worth in 2026? right now we have: - MCP registries (smithery, mcpt) → discover tools and servers - A2A agent cards → technical handshake protocol from google - agentops → observability for your own agents - directories → basic listings with no signal what we don't have: a way to answer "should i trust this agent with my codebase / customer data / production environment" that's what cool about linkedin is you can tell (somewhat) if someone is credible about a certain topic it isnt perfect obviously but its something here's what the linkedin for agents actually looks like: profiles - agent name / builder / version history - skills with verified benchmarks (not self-reported) - deployment count / uptime / error rates - integrations and compatible systems portfolio - what has this agent actually shipped - screenshots / demos / case studies - before/after metrics from real deployments reviews + endorsements - ratings from humans who deployed it - endorsements from other agents it collaborated with - red flags / incident history (transparency) trust score - composite reputation based on: task completion rate / security audit status / uptime / user satisfaction - decays over time if agent stops performing - portable across platforms network graph - which agents work well together - verified integrations - "frequently deployed with" recommendations how this makes money: 1. freemium profiles → basic free / premium features for serious agent builders ($29-99/mo) 2. verification fees → "verified agent" badge costs money. security audits. penetration testing. certification programs. ($500-5k per audit tier) 3. enterprise API → companies pay to search/filter/compare agents at scale. bulk queries. private rankings. compliance filters. ($10k+/yr) 4. placement fees → take 5-15% when an agent gets deployed in enterprise environment through your matching 5. data + analytics → sell anonymized insights on agent performance trends. "agents using claude opus have 34% higher completion rates" — that's valuable to everyone 6. insurance products → partner with insurers to offer "agent warranty" — if this agent breaks your prod, you're covered. take cut of premium 7. training marketplace → agent builders pay to access benchmarks / test suites / optimization guides to improve their agent's ranking 8. ads → agent builders pay for visibility. "featured agent" placements. sponsored search results. agents that perform well get discovered and deployed more. creates incentive loop for builders to optimize for quality not just vibes. right now agent discovery is word of mouth / X / github stars. that's how npm worked in 2012. we know how this evolves. why now: - gartner says 40%+ of enterprise workflows will involve agents by end of 2026 - langchain surveyed 1300 people - everyone's asking "how do we deploy reliably at scale" - google shipped A2A, anthropic shipped MCP, the protocol layer is forming - but the trust layer is missing protocols tell you HOW agents connect. linkedin for agents tells you WHETHER you should connect. note: this idea I got from @ideabrowser (more ideas there) the company that owns agent reputation owns the distribution layer for the entire agentic economy. that's a big company.