@ryanjunee

Building superintelligence alongside meditative practice to discover the nature of mind.

 here
ryan.xyz
Joined March 2007
  • Tweets3.5K
  • Following974
  • Followers5.9K
271Media
Include

Only show posts containing:

Exclude

Hide posts containing:

Time range
-
Minimum likes
Replying to @LunarResearcher
the checks have to live outside the agent's reach or it optimizes against them. Mine sit in CI as product invariants the loop can't edit. xitter.cf/ryanjunee/status/20807…
Ryan
@ryanjuneeJul 24
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
22
Replying to @hanakoxbt
Judge drift is the underrated one on this list. I eventually gave up and swapped the judge for a deterministic grounding check. Over-blocks sometimes but it can't drift. xitter.cf/ryanjunee/status/20801…
Ryan
@ryanjuneeJul 23
Is it possible to build enterprise-grade AI agents that don't hallucinate? I've been tackling this problem and have come up with a potential solution that uses a deterministic gate at the end of each turn, to trace all facts mentioned in a response back to the original source data. It comes with trade offs, but this is the type of thing that can enable more widespread AI usage in even the toughest regulated industries.
45
Replying to @asklivermore
Everyone bought the model first, then found out accuracy is its own engineering project. wrote up how I handle the fabrication half xitter.cf/ryanjunee/status/20801…
Ryan
@ryanjuneeJul 23
Is it possible to build enterprise-grade AI agents that don't hallucinate? I've been tackling this problem and have come up with a potential solution that uses a deterministic gate at the end of each turn, to trace all facts mentioned in a response back to the original source data. It comes with trade offs, but this is the type of thing that can enable more widespread AI usage in even the toughest regulated industries.
24
Replying to @freeCodeCamp
One solution is a a deterministic gate that traces every fact back to data fetched that turn, here's the version I run xitter.cf/ryanjunee/status/20801…
Ryan
@ryanjuneeJul 23
Is it possible to build enterprise-grade AI agents that don't hallucinate? I've been tackling this problem and have come up with a potential solution that uses a deterministic gate at the end of each turn, to trace all facts mentioned in a response back to the original source data. It comes with trade offs, but this is the type of thing that can enable more widespread AI usage in even the toughest regulated industries.
5
Replying to @prasad_pilla @bcherny
Nice, I wrote something similar opening with that exact same Cherny quote :) xitter.cf/ryanjunee/status/20807…
Ryan
@ryanjuneeJul 24
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
64
Replying to @0xJeyx
Graph buys you parallelism. Keeping it honest deterministic gates. here are some notes on how I do it xitter.cf/ryanjunee/status/20807…
Ryan
@ryanjuneeJul 24
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
1
1
59
Replying to @biektive
I run roughly that shape per task (plan, build plus review, then a synthesis pass at the end). xitter.cf/ryanjunee/status/20807…
Ryan
@ryanjuneeJul 24
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
1
47
Replying to @0xCodila
Grounding the evaluator helps but it doesn't stop two agents nodding along to each other's mistakes. Telling the reviewer to refute rather than assess is what finally fixed that for me. xitter.cf/ryanjunee/status/20807…
Ryan
@ryanjuneeJul 24
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
2
597
Replying to @noisyb0y1
The multiplier is real but it points both ways. Skip the gate in that Evaluate step and it multiplies the confidently wrong output just as fast
35
Replying to @plutos_eth
Yeah, and it needs a bound too. I cap the builder/breaker argument at 2 rounds then escalate, the unbounded version just burns tokens arguing with itself. xitter.cf/ryanjunee/status/20807…
Ryan
@ryanjuneeJul 24
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
58
Replying to @Awesome_O_AI
Stopping criteria took me longer than everything else combined. Green tests alone weren't enough, it needed a spec-drift check and product invariants in CI before I could walk away from it. here's what that ended up looking like xitter.cf/ryanjunee/status/20807…
Ryan
@ryanjuneeJul 24
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
58
Replying to @alex_prompter
In my case, every node in the graph turned out to be a loop itself, so the motor runs inside the rails rather than next to them. wrote up where I landed xitter.cf/ryanjunee/status/20807…
Ryan
@ryanjuneeJul 24
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
164
Graphs? Loops? How about a graph of loops! I don't know why it needs a name tbh, but this is what I've been running for months. Runs autonomously for days or weeks, pulling me in occasionally to answer questions that it can't resolve on its own.

A Graph of Loops

Article · @ryanjunee
1
1
11
2,347
neutronenterprise.com/insigh…

An Enterprise-Grade Architecture for AI That Doesn't Hallucinate - Neutron Enterprise

How to make hallucination structurally impossible in enterprise AI: a deterministic grounding gate, turn-scoped fact corpora, identity anchoring, configurable strictness, and the real bugs we shipped...

neutronenterprise.com
48
Is it possible to build enterprise-grade AI agents that don't hallucinate? I've been tackling this problem and have come up with a potential solution that uses a deterministic gate at the end of each turn, to trace all facts mentioned in a response back to the original source data. It comes with trade offs, but this is the type of thing that can enable more widespread AI usage in even the toughest regulated industries.
1
1
2
281
Replying to @AdamGillman
X
25
Load more