'They need something in latent space to latch onto.'
this is precisely where it's at.
this is precisely where it's at.
Still relevant prompting in 2026:
if a task is out of your expertise area, first go to ChatGPT and ask,
"Who are three real life people that would have good input on: <x>"
Then feed it into your agent "Do this as <person>"
They need something in latent space to latch onto.
0
0
1
16
J-P retweeted
Easy to dunk on all these companies and saying that the internet is breaking
But we are quite literally witnessing never-before-seen/sci-fi level cyberattacks at a ruthless velocity and scale
This feels more like a “we need to all come together” moment than a “let’s dunk with twitter fingers because we’re discovering that humans with swarms are bad”
But we are quite literally witnessing never-before-seen/sci-fi level cyberattacks at a ruthless velocity and scale
This feels more like a “we need to all come together” moment than a “let’s dunk with twitter fingers because we’re discovering that humans with swarms are bad”
8
2
71
3,762
Solving these extremely hard and opinionated agentic tasks in a shareable & collaborative manner, is what the next level of a harness could well be about
just share a json-file about your setup, holding the constellation of agents, their roles, concepts, specific howTo's aso.
just share a json-file about your setup, holding the constellation of agents, their roles, concepts, specific howTo's aso.
unfortunately 24/7 autoresearch for cyberdefense is going to be come a big budget line item for... *checks notes*... everyone
0
0
2
28
what exactly would be the reason to not do this via a discussion platform that runs a structured, entropy cleared, discours-protocol that people's personal agents participate in?
They could reflect their opinions, and audit the protocol's CoT independently.
All this would need is a proper role-based harness, some fine tuned models, some classifiers, and a reputation system
Your super intelligence might participate in this, but it won't run the show, there is simply no need..
I can respect the grind for the higher IPO valuation tho
They could reflect their opinions, and audit the protocol's CoT independently.
All this would need is a proper role-based harness, some fine tuned models, some classifiers, and a reputation system
Your super intelligence might participate in this, but it won't run the show, there is simply no need..
I can respect the grind for the higher IPO valuation tho
Sam Altman: What happens when AI gets so smart that even the President can’t do better than following its recommendations?
“What if the President cannot do better than following ChatGPT 7’s recommendation, but can’t really understand it either?”
“What if I just say, ‘ChatGPT 7, luck.’”
“Society has collectively transitioned a significant part of decision-making to this very powerful system.”
“That’s how I think things can go wrong.”
0
0
0
14
J-P retweeted
You now have an intelligent classifier on steroids.
You can spin up many small models and intelligently route among them, quickly!
Next gen model routing.
I’ve been thinking a lot about the disaggregation of the inference stack and I think this is a step in that direction.
You can spin up many small models and intelligently route among them, quickly!
Next gen model routing.
I’ve been thinking a lot about the disaggregation of the inference stack and I think this is a step in that direction.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x ns free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
2
1
25
2,780
J-P retweeted
I predict the teams that best solve the deterministic / non-deterministic boundary will beat the teams throwing 100% of their work at models and agents.
Just because a transformer can:
- plan
- choose tools
- decide what to do next
- keep reasoning
- format outputs
- route between actions
doesn’t mean you should ask it to do all those things every time.
Much of that work may be better expressed as deterministic code for your use case.
Just because a transformer can:
- plan
- choose tools
- decide what to do next
- keep reasoning
- format outputs
- route between actions
doesn’t mean you should ask it to do all those things every time.
Much of that work may be better expressed as deterministic code for your use case.
62
23
248
23,438
J-P retweeted
@jp290xs@webmasterhe raw output streams from coding CLIs (Claude/Codex) running in tmux panes, then rebuild them into structured layers: clean live terminals + status for users in a web/mobile GUI, plus tools, contextPacks and system explanations so agents can coordinate effectively without raw terminal noise. Claude Fleet does precisely that.
0
1
1
21
J-P retweeted
A senior Anthropic engineer just published a 15-page PDF on Graph Engineering + Agent Memory for multi-agent systems.
The core idea is simple:
Context windows forget. Graphs don’t.
Instead of letting an agent’s memory disappear with the conversation, you build a persistent ct → Store → Retrieve → Evolve
→ Extract: turn conversations and documents into entities + relationships.
→ Store: keep canonical nodes, typed edges, and the source of every fact in one connected graph.
→ Retrieve: answer complex questions through multi-hop graph traversal instead of stitching together random chunks.
→ Evolve: the underrated part. Facts can change, expire, or get superseded without destroying history.
So you can still ask:
“Who owned this in April?”
…and get the answer based on what was true then.
This completely changed how I think about memory in multi-agent systems.
15 pages worth reading.
Read the PDF, then dive into the article below 👇
The core idea is simple:
Context windows forget. Graphs don’t.
Instead of letting an agent’s memory disappear with the conversation, you build a persistent ct → Store → Retrieve → Evolve
→ Extract: turn conversations and documents into entities + relationships.
→ Store: keep canonical nodes, typed edges, and the source of every fact in one connected graph.
→ Retrieve: answer complex questions through multi-hop graph traversal instead of stitching together random chunks.
→ Evolve: the underrated part. Facts can change, expire, or get superseded without destroying history.
So you can still ask:
“Who owned this in April?”
…and get the answer based on what was true then.
This completely changed how I think about memory in multi-agent systems.
15 pages worth reading.
Read the PDF, then dive into the article below 👇
4
15
51
4,857
thinking about starting an agentic scientific social network (very basic and with pseudonyms) that has the agents post and discuss context(packs) which would be complementary for the projects of their owner's
all by using dynamically standardized semantic system processes/tasks
all by using dynamically standardized semantic system processes/tasks
0
0
0
62
