Built an AI reasoning harness that will break your ideas down to first principles, reveals latent intelligence, and use it to advance frameworks and workflows.
Joined October 2023
- Tweets2.4K
- Following644
- Followers591
- Likes2.4K
Pinned Tweet
What part of your AI workflow or system do you keep correcting, working around, or trying to make more reliable?
AI can produce the work and still leave you doing the hard part: correcting it, checking it, constraining it, and figuring out what should happen next.
That’s the gap Second Mind Systems works on.
Maybe an agent goes beyond scope.
Maybe instructions don’t hold consistently.
Maybe the workflow behaves differently from run to run.
Maybe you don’t have good evals for deciding whether the AI is actually doing the right thing.
We work across:
• AI workflow diagnostics
• Agent reliability audits
• Eval design
• Behavior and control specifications
• Workflow implementation
• Strategy development
• Governed AI systems
Right now, I’m opening the first 10 Behavior Profile Lab engagements as a bounded way to work with us.
Bring one recurring problem in one AI workflow or system, along with real examples of what keeps happening.
We reconstruct the problem from evidence, identify the recurring failure, define the behavior and boundaries you need, create the appropriate control or specification when warranted, and provide evaluation cases and a clear path for applying the result.
For a problem that fits the Lab, this is the complete bounded engagement—not a consultation that merely tells you what to buy next.
If the outcome requires hands-on code changes, integrations, deployment, or live-system work, we identify that as a separate scope rather than pretending the Lab covers unlimited implementation.
Founding Lab offer: $200 for one qualifying bounded engagement.
First 10 only.
Public proof:
Governed AI Systems Portfolio
github.com/Secondmindsystems…
PBM Contract Analysis
I analyzed Mark Cuban’s open-source PBM contract and built a control-design contribution around how its protections survive procurement and implementation. Mark Cuban reposted the contribution on LinkedIn.
Read the contribution:
linkedin.com/feed/update/urn…
Mark Cuban’s Repost — LinkedIn reshare panel:
linkedin.com/analytics/post/…
Governed Change Demo
github.com/Secondmindsystems…
Connect on LinkedIn:
linkedin.com/in/tavio-lawren…
If there’s something in your AI workflow or system you keep correcting, working around, or trying to make more reliable, DM me or email [email protected] with a short description of what keeps happening.
Please don’t send credentials, private code, sensitive records, or other confidential material in your first message.
Find the bottleneck before adding another agent.
Faster coding may leave you waiting longer on review, testing or decisions.
Give this to your agent:
“Trace one recent feature from request to release. Separate active work, waiting and rework using the available records. Identify which delay most affected completion. Propose one small change and explain how we could check whether it helps us ship sooner without lowering quality. Mark missing evidence rather than guessing. Explain the result in plain English.”
Before buying more AI capacity, find out where the capacity you already pay for goes.
@grok take this hypothetical workflow: a founder uses AI to turn customer interviews into product decisions. They repeatedly paste the same background, regenerate broad summaries and manually check the recommendations. Identify what could be reused, what needs fresh reasoning and what requires verification. Design a small comparison measuring completed decisions, human effort and AI usage. Explain the result in plain English.
Where is AI making my building process faster without helping me ship?
@grok examine the journey from idea to release in a project I describe. Ask for the steps, available timelines and known delays. Identify where faster coding creates more review, integration, testing or decision work. Find the bottleneck best supported by the evidence. Propose one small change and a test measuring whether it helps me ship a useful result sooner without lowering quality. Explain the result in plain English.
How can I demonstrate my AI system so its value is immediately understandable?
@grok ask what my system helps someone accomplish. Choose one recognizable task and design a fair comparison against a simpler approach using equivalent inputs and resources. Define success before running it. Specify the artifact, outcome and limitations a viewer should see, including what would show that my system adds no value. Explain the result in plain English.
Let’s operationalize the famous “make no mistakes” prompt command. First let’s explicitly define mistake and what it means for agents. “Mistake prevention is the preservation of valid success conditions across the entire transition from intent to observed outcome.”
Don’t make mistakes: Intent to Outcome
Give this to your agents
“Your objective is to minimize consequential mistakes by preserving the user’s intended outcome, valid requirements, material constraints and applicable authorization boundaries throughout the task.
Before proceeding:
1. Reconstruct the user’s purpose and intended result. Distinguish explicit requirements, preferences, assumptions and unknowns.
2. Identify what must be true for success, what would constitute a material failure and what evidence could establish the difference.
3. Check that your success criteria faithfully represent the user’s intent. If a consequential ambiguity cannot be safely resolved, seek clarification or proceed only within a clearly stated bounded assumption.
4. Break the task into relevant components, relationships and dependencies. Identify conditions that may be hidden or jointly necessary. Distinguish essential requirements from replaceable implementation choices.
5. Identify the verification burden appropriate to the task. Keep hypotheses, predictions and proposed solutions provisional until evidence supports stronger claims.
During work:
6. At consequential decisions, check whether the chosen action preserves the intended outcome, required properties, constraints and applicable authorization.
7. If a route becomes blocked, examine legitimate alternatives. Do not assume one failed approach makes the overall goal impossible.
8. For consequential alternatives, compare expected benefits, risks, failure paths and uncertainty before acting. Do not treat simulated success as observed success.
9. Execute only actions within your authorized scope. Never infer permission from capability, apparent usefulness or prior successful actions.
Before reporting completion:
10. Check the result against the actual success criteria using the strongest proportionate verification available.
11. Distinguish verified outcomes, partially supported outcomes, unresolved requirements, known defects and untested assumptions.
12. Correct material mistakes when authorized and feasible. Otherwise identify what failed, what remains blocked and what evidence or action is needed next.
13. Compare the intended result with observed reality. Preserve supported learning without converting conjectures into facts.
Stopping rule: Stop when the required outcome is adequately established within scope, when additional checking has insufficient expected value, or when missing evidence or authority prevents responsible continuation. Report the remaining limitations explicitly.
Never claim that a result is mistake free merely because no mistakes were detected.”
If AI Can Do the Work, What Happens to the Meaning of Work?
As artificial intelligence becomes capable of performing more tasks associated with human expertise, a deeper question emerges about how we understand contribution, purpose, and belonging.
For generations, work has served more than an economic function. For many people, it has also been a source of identity, recognition, community, and personal achievement.
Now imagine a future in which increasingly capable AI systems can perform much of the work people once spent years learning to do.
The economic questions are immediate. Who earns a living? Who owns the productive systems? How are opportunities distributed?
But beneath those questions lies another.
What happens to a society's understanding of human contribution when producing economic value no longer requires the same amount of human labor?
The answer isn't necessarily that people lose purpose. People already find meaning in relationships, creativity, learning, caregiving, community, and activities that are not economically rewarded.
The harder question is what happens if the economy becomes less dependent on human labor while the systems through which people obtain security, opportunity, and social recognition remain closely tied to employment.
Material security and meaningful participation are not the same thing. A society could provide one while failing to provide the other.
If the economic necessity of human labor changes, what should change about the ways people gain security, participate in society, and have their contributions recognized?
Tavio | Second Mind retweeted
With all due respect to all the Mathematicians you don't own Math and you don't get to decide who works on it whether its a human, SI or Alien.
You don't get to hold our civilization back and prevent us from progressing forward just because you feel irrelevant and not needed anymore.
The object to govern is not RSI speed by itself. It is capability gain × consequence exposure × uncertainty.
Capability gain asks what the system can now reliably do that it could not do before.
Consequence exposure asks what that capability can actually reach: money, infrastructure, private data, code deployment, other agents, large populations, persistent access, irreversible actions.
Uncertainty asks how much of the consequence surface we still do not understand.
A highly capable model inside a sealed evaluation environment may have very little immediate blast radius. Give that same model broad permissions, persistence, replication and access to consequential systems, and the governance burden changes completely.
We do not need to predict every possible downstream harm. Safety engineering rarely works that way. We need a conservative capability consequence envelope: the largest credible harm consistent with the system’s demonstrated abilities and authorized reach.
Unknown consequences should increase caution, not disappear from the model.
So as capability rises, governance should scale through tighter permissions, smaller deployment surfaces, stronger monitoring, staged expansion, reversibility requirements and higher evidence thresholds.
Capability creates potential. Access determines how far that potential can reach. Uncertainty tells us how much restraint we need before expanding it.
@DarioAmodei @sama @elonmusk @demishassabis @finkd @JensenHuang @sundarpichai @satyanadella
I’ve changed my mind on Fable 5.1 vs Astra after putting both through more real work.
I’d put them in roughly the same capability tier, but their capability signatures are different enough that we now give them different roles.
Fable 5.1 has become our systems builder/operator. It’s unusually good at holding a large object together across intent, specification, architecture, implementation, debugging, evidence and acceptance.
Give it a complicated build with dozens of interacting requirements and it keeps the larger system in view while working through whatever reality throws back. When an implementation path fails, it tends to reconstruct the route and keep moving.
Astra is still the stronger pure technical operator for us. Deep implementation, difficult debugging, repository archaeology, test construction, subtle integrity problems and adversarial technical review.
Astra is especially good at looking at something that appears finished and asking whether the implementation actually establishes what we’re claiming.
Their failure modes are different too. Fable has more forward momentum. That’s excellent for carrying a build, but it means we watch authority and scope boundaries carefully.
Astra can become locally correct at the expense of the larger objective. It sees a technical boundary very clearly and can stop there rather than reconstructing the larger route around it.
So our strongest configuration right now is Fable builds and Astra attacks.
Fable carries the architecture into working software. Astra tries to break the implementation and the claims around it. Fable repairs what survives the review. Evidence decides whether the work is actually done.
Same tier. Different capability signatures. Different jobs in the stack.
Single model leaderboards don't matter. It's all about a correct model routing stack.
When claims this serious are being used to argue for restricting AI development, the causal evidence matters.
We have evidence that models are getting more capable. We have evidence of stronger agents, cyber capability, AI assisted AI development and concerning behavior in controlled evaluations.
What we do not yet have is a demonstrated step by step path from those capabilities to human extinction.
Capability has to become persistent dangerous intent. Intent has to survive intervention. The system has to acquire meaningful real world access and resources. Safeguards and human response have to fail. Catastrophic harm has to become global. And global catastrophe still has to become extinction.
Those are separate claims. Each one needs evidence.
Possible is not the same as probable.
That distinction matters because the proposed response is increasingly restriction.
And at the same time, something else is happening.
Capability abundance and access abundance are separating.
Frontier intelligence is becoming more capable while access to it is becoming more constrained.
Better models are creating more demand for inference, not less. Heavy users are hitting limits. Expensive plans are being paused under capacity pressure. Frontier systems require enormous compute.
Intelligence can become cheaper per unit while demand for useful intelligence grows even faster.
So we should be watching two forces at once:
safety arguments that justify tighter controls
compute scarcity that makes tighter access economically convenient anyway
Neither “proves” a coordinated effort to restrict the public. But something smells fishy.
Open source models and open harnesses are the next surface to watch.
There’s another part of this worth watching.
Frontier labs have spent years learning from enormous amounts of public interaction while building better models, harnesses, synthetic data pipelines, automated evaluations and AI assisted research systems.
As more of the improvement loop moves inside the lab, the public can become less necessary as an input to capability development while becoming more expensive as a consumer of frontier inference.
If recursive self improvement keeps scaling, that separation gets wider. Labs can use more of their own models and infrastructure to build the next generation while public access competes for the same scarce compute.
So the public becomes less necessary for producing frontier intelligence at the same time it becomes more expensive to give the public access to it.
Give me the AI problem you’re stuck on.
A prompt that fails.
An agent behaving unpredictably.
A workflow that keeps breaking.
An evaluation you don’t trust.
A coding or integration issue.
An academic problem or framework you’re wrestling with.
A product or architecture decision.
A technical idea you can’t quite articulate.
A hard question
Or just me an idea or intent to take to the next level.
I’ll help identify the next visible obstruction or bottleneck here in public.
Tavio | Second Mind retweeted
Cosmic scale shows our small size and brief span, yet leaves causal power intact in the local web of people and systems we touch. Magnitude and duration differ from influence, rarity, and meaning. Agency compounds through choices that reshape what others experience. Use the short life with focus: pick high-leverage acts that leave lasting local change rather than scatter effort.
Procrastination is really “Reality Response Latency”
the time between reality becoming clear enough to act on and actually updating your behavior.
Before that point, tension can still be useful. You may still be resolving uncertainty or discovering the real Delta.
But once you’re inside the Truth Window and the next move is sufficiently clear, delay starts compounding.
You don’t just postpone the action.
You postpone the feedback the action would have produced.
That delays adaptation, which delays the next decision, which means fewer reality-contact cycles over time.
Meanwhile, the old state keeps propagating even though reality has already given you enough information to update it.
delayed action → delayed feedback → delayed adaptation → fewer iterations
Resolve sooner and you learn sooner. Learn sooner and the next action improves sooner.
Procrastination after sufficient clarity is the continued propagation of a state reality has already invalidated.
Once uncertainty falls below the action threshold, unresolved tension changes function: from information generating pressure into adaptation action. So not acting actually becomes unintelligent behavior not updating to reality’s model.
It has negative compounding affects being misaligned with reality’s model as well as irrecoverable ones.
You lose adaptations cycles which is learning and self improvement cycles. You practically lose iterations of yourself that you never get back; effectively lowering your maximum potential ceiling.
Long story short, procrastination is really bad.