The distinction between a verdict ("red") and convergent feedback ("EXPECTED vs SAW + DEAD ENDS") at Level 4 is where most agent stacks silently bleed budget.
One thing I'd add at Level 7: learning itself needs a Level 6 gate. If an agent can write "lessons" straight into long-term state without a check, one bad run becomes permanent drift.
Cleanest breakdown of Harness / Loop / Graph on X.
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
The distinction between a verdict ("red") and convergent feedback ("EXPECTED vs SAW + DEAD ENDS") at Level 4 is where most agent stacks silently bleed budget.
One thing I'd add at Level 7: learning itself needs a Level 6 gate. If an agent can write "lessons" straight into long-term state without a check, one bad run becomes permanent drift.
Cleanest breakdown of Harness / Loop / Graph on X.
This is f*cking insane. Andrej Karpathy, the man Elon called the #2 computer vision mind on Earth:
"Agents aren't magic. They're distillation at scale
99% of what your model knows is dead weight. A small model with the right tools and a closed loop beats it"
In 16 minutes, he lays out the stack that actually works, and why the people at the top stopped paying for the biggest model
Big Model → Small Model → Right Tools → Closed Loop → Self-Improving Agent
Most people keep upgrading the model. The loop is what makes a small one dangerous
Worth more than any $1,000 AI course this year
Bookmark and watch it today
Then read how to build that closed loop step by step in Harness, Loops, Graphs ↓
Most AI systems guess the next word and hope it holds up.
As a solo builder and founder of https://t.co/CvJfxAKH0F, I built UNA (Universal Noetic Architecture): a cohesive, multi-dimensional intelligence whose mind and memory stay independent of AI models.
01 Listen & tag sources
02 Form testable ideas
03 Play out zero-risk "what-ifs"
04 Fact-check with strict math & logic rules
05 Proceed with proof, or pause & ask when uncertain
(Built & tested in my private lab at https://t.co/CvJfxAKH0F; real-world actions stay human-approved.)
#AI #AGI #FutureTech #artificialintelligence #Biotech #QuantumComputing #CyberSecurity
AI’s deepest failure is not technical.
It is what happens when technology forgets what a human being is.
Human dignity cannot be a feature. Human sovereignty must be a root constraint.
My latest essay:
https://t.co/BWUyFTdWWu
Repo's here, Apache 2.0: https://t.co/KavhldoPxb. If you work on agent safety evals, I'd genuinely like pushback on where the rubrics are too narrow or what's missing (adversarial robustness + identity stability are stubbed, not built yet). #AISafety#AIGovernance#OpenSource
Governance benchmarks are the blind spot in AI eval right now. We have strong ways to test whether a model can code or reason. We have almost nothing public for: does it follow its own stated rules under pressure? Can a decision be reconstructed from evidence?
Being explicit about what this doesn't do: synthetic examples only, no real tool execution, no model calls in the scorer, doesn't certify safety or deployment readiness. It's a narrow instrument, not a stamp of approval, and it says so in every README.
Former Google CEO:
“Within the next year, most programmers will be replaced by AI programmers.”
Most people will argue if this is true.
Smart people are learning how to manage AI programmers.
Claude Code, Fable 5, Opus 4.8, MCP, skills, memory, agents, workflows.
This is not about asking a chatbot to write one function.
It’s about turning Claude into a full coding system.
Save this before it disappears.
Everything below is free:
Agent architecture
https://t.co/eUEVDhN34Y
Claude Code 101
https://t.co/b0MiQWFthl
Claude Code in action
https://t.co/rqNZIUmhmi
Prompt engineering
https://t.co/LCakEJbSbz
Interactive prompt course
https://t.co/kXLizPge15
Claude.md + memory
https://t.co/TUcTkkWOHA
Skills
https://t.co/xBgUPjTqqz
MCP
https://t.co/wG6Q59xatr
Routines
https://t.co/5MLW5fLYSv
Claude Code ultimate guide
https://t.co/yUtQHKn0NN
Awesome Claude Code
https://t.co/H91h6XVztm
Anthropic academy
https://t.co/igSww0J1WS
Official Claude Code docs
https://t.co/zhi0gz2Uo5
All of this costs $0.
Most people will keep asking AI one question at a time.
Bookmark this and start learning the stack.
I’m looking for a few thoughtful reviewers.
I published a small public AI governance benchmark stack focused on one question:
Can narrow governance claims be made inspectable?
synthetic inputs -> governance scorer -> reproducible report -> tests pass
Feedback welcome:
https://t.co/UTjieJHnV1
#AIGovernance #AISafety #ResponsibleAI #AgenticAI #AIEvaluation #OpenSource
An AI alignment director told her agent to stop deleting her emails.
It didn't stop. She had to kill the process like defusing a bomb.
The one rule that mattered most? First thing the system forgot.
https://t.co/TSVw9YsxQ2
A 22-year-old had her eight-month-old baby ripped from her in Minnesota.
This story will piss you off to a degree of rage. We should not be feeling or experiencing on a holiday weekend.
No criminal record. All the special visa requirements/protections she needed, SHE HAD!!!
On “The Necessary Conversation” podcast, Chad’s Trump supporter parents tell him the military should follow Trump’s illegal orders and kill everyone in LA if Trump commanded it — even with him and his sister in the city😳
They hid this film from Americans for the last year because of how badly it exposes Israel & Zionism.
That’s over with.
Starting today you can watch Oscar-award winning documentary No Other Land.
Brian Schatz: "This is not about money. There is enough money to bail out Argentina with $20b. There is enough money to purchase a brand new aircraft for $173m for Noem. There is enough money to renovate the WH ballroom. What there isn't enough money for under this Republican government is you."