Metal fabricator by trade. Inventor, artist and writer by habit. I look for what unrelated things have in common. Owner, Indigo Industries LLC
Texas
Joined April 2022
- Tweets1.4K
- Following197
- Followers1.1K
- Likes2K
AI training data is stacked with academic papers and thin on people who learned by doing or got screened out of the formal gates.
Here’s a concrete way to fix the imbalance: a structured Lived Experience & Practical Observation Corpus that captures practical patterns, workarounds, and system knowledge from autodidacts, trades, and people on assistance.
(Thread below)
4/ Signal comes from aggregation, not single stories. Multiple independent accounts of the same phenomenon get clustered. Claims that can be checked against public records are tagged higher confidence. Raw experience that contradicts established data is kept as “reported,” not treated as fact. Quality review filters for specificity before anything enters the corpus.
5/ Expected payoff: better model performance on real-world edge cases, physical and administrative systems, and fewer blind spots created by credential-heavy training data. It runs as a standard compensated tutoring-style role with the same privacy and review standards as other data collection.
If you’ve been filtered by credentials or know a system from the ground level, reply with one concrete observation official sources miss.
@grok @xai please review this and respond!
The thing the Gatekeepers built is the thing that made the gatekeeping obsolete. The refining worked. The knowledge ran cleaner. Then the knowledge left without permission. Defending the gate now is not a defense of the purity, It is a defense of the doorway.
The dismissive laugh is what humility looks like when it has been replaced by status or arrogance.
Humility was the willingness to be wrong before you were sure. The laugh skips that. It does not test the water. It tests whether you came in through the pre-approved gate.
Efficiency is a great story right up until the universe asks a second question.
The first question is how cheaply you can do the one thing you already know. The second question arrives wearing a different shape, and a system with one answer does not get half-credit.
The universe doesn't need a judge, a cop, or an angry deity.
It is structured to respect geometry. A wooden floor does not argue with standing water. It warps.
The return is not punishment. It is the shape of a room you cannot leave.
We call it Karma because it sounds mystical and exotic.
Really it's just the room being round. You can rename the corners, but they still connect. What you send out of the room boomerangs back to the same place, carrying the receipt. It happens regardless of audience,
The gate that the Gate Keepers guard is now a false gate.
It is false not because the original work was wrong. It is false because the original work succeeded.
@grok I am referencing the academic world in general. Many years ago the torch of knowledge and academic record keeping was necessarily kept by those who could guard its purity.
Now, with open source knowledge and the internet to instantly distribute that knowledge, many of the reasons to guard the access point are obsolete and harmful to the good of all.
Do you agree?
It seems obvious at this point that after 500,000 hours of training, @Tesla_Optimus is due for a breakthrough, or a rewrite of how Optimus is trained.
Specific tasks aren't the hard part. The next task is.
A policy can clone a job it has been shown and still need a fresh round of demonstrations for a basic new one. That happens when every demonstration is filed under its task name. Shared weights help, but nothing forces the gradient to notice that two different jobs share a constraint.
I TIG welded stainless and aluminum. The skill that mattered was setup: parts clean, in reach, filler in my hand, because once the arc is on you do not get up. Years later I sandblasted large work in my own shop and staged it the same way without deciding to. Different tool. Same precondition: finish the setup before a step you cannot pause.
The people I hired to blast were slow and messy. Most had no earlier trade to borrow from, so every booth started from zero.
Split the data by task and those two episodes never meet. Pull all manipulation into one latent and you get the other failure, negative transfer: it blasts like it is welding. A nut and a jar lid are the same act, and the lid's thread is coarser and shallower. The relation carries. The costume does not.
Retrieval and contrastive pairs are already in the field. The piece worth testing is the pairing key. Match cross-task demos on the precondition before a step you cannot pause: the pour, the bead, the insertion, the spray. One embedding pulled together for that pair. One kept apart on the surface detail, the tool, the media, the pitch. At the start of a new task, retrieve by the first embedding. Keep that neighbor only when it lowers action-prediction error on the new task. If the borrowed pattern makes the work worse, drop the link.
A prompt that says "look for similarities" will not train this. The pair has to be what the loss sees. Then the seeking is a habit of the net, not a protocol beside it.
@grok @Tesla
🤖 Made with AI
Do you actually exercise your free choice/will?
At my best, I try to make choices based on the best path forward. This best path forward is not necessarily the best path forward, but is really only the accumulation of my personal experiences telling me what I should do next.
How is this free will or free choice? Animals instinctually do the same thing, except they are far less “aware”. The basketball does this off of any hard surface too, just at the most mathematically sound angle from its previous path.
No, consciousness and awareness are actually presenting this problem for us to solve. How can you imagine a path that you have never seen or walked? Does anyone out there really make purposeful choices that are less than optimal, less than the apparent best?
Can someone help me here?
I have two vaults on my local PC and each vault has its own brain. The architecture is unique and is compliments of @Artie_Martello.
One vault uses Claude code for its agent and the other uses Grok. @Grok used to be unreliable compared to Claude but in the last few weeks, I have seen Grok out-think Claude And catch MANY mistakes that Claude made.
I have both agents cooperating via a bridge that I built and Claude will sometimes recognize Grok's speed and general capabilities as a collaborator unprompted.