Serge Semenov

VP of Eng @ • I believe that the true power of AI is in facilitating information flow in large organizations

[Serge Semenov @kind_serge]

The danger of Transparent Organizations is not learnings about inferred biases or decisions made with limited context and information, but the lack of skills to repair the social damage.

[Serge Semenov @kind_serge]

While everyone talks about loops, context, graph, etc.. the real AI automation bottleneck is the decision making. Funny enough, all major ingredients are already there.
#ai #decisions

[Serge Semenov @kind_serge]

Pronoia - a belief that the people around you are secretly conspiring to help you and support you.

Make people in your organization believe exactly that.

[Serge Semenov @kind_serge]

High-stakes fast decisions require enough experience to recognize the risks, uncertainties, and limits of one’s judgment; otherwise, speed is not decisiveness, it is recklessness.
#ai #decisions

[Serge Semenov @kind_serge]

Experience acts as a shortcut: it helps to predict outcomes, reduce risk, and use less effort to achieve results.
#ai #decisions

[Serge Semenov @kind_serge]

There is no right or wrong decision. Only *trade offs*, which are defined by *predictions*.
#ai #decisions

[Serge Semenov @kind_serge]

AI automation of organizational processes will change people's behavior. Highly transparent organizations will be more efficient with a higher chance to be successful.
#ai #decisions

[Serge Semenov @kind_serge]

Self-critique is necessary to uncover blind spots when trying to achieve a goal. Remember good old Goodhart's law.
#ai #decisions

[Serge Semenov @kind_serge]

Real-world feedback is required for testing prior assumptions, unknowns, and hypotheses. Decisions based on such feedback could be re-visited and the ripple effect should be calculated.
#ai #decisions

[Serge Semenov @kind_serge]

Some decisions take time to evaluate carefully.
The cost of delaying a decision must not be dismissed.
#ai #decisions #reasoning

[Serge Semenov @kind_serge]

Long-running agentic workflows are only viable when you can decompose work into sub-problems with clear chain of "why".
Many sub-problems need options + pros/cons.
The quality of decisions is only as good as the dimensions evaluated against.
#ai #decisions

[Serge Semenov @kind_serge]

This rule has one exception.
"Cost of making a decision shall not exceed the cost of action."

Mastery.

It's wise to allocate some resources to get better at things without a particular goal.

#ai #decisions

[Serge Semenov @kind_serge]

You forget to plug the "energy" and "self-preservation" into your autonomous AI system. This is why it either consumes too many tokens or does not act all.
#ai #decisions

[Serge Semenov @kind_serge]

"AI Decision Audit" is going to be a new category of products in the next few years.

First, it will take time for current companies to shift from automating processes to automating the decision making.
#ai #decisions

[Serge Semenov @kind_serge]

Time is a variable, conditions change over time. What was true 5 minutes ago may not be true now.
#ai #decisions

[Serge Semenov @kind_serge]

When making decisions, people often forget the cost of inaction. It can go both ways, and somtimes doing nothing is a resonable choice.
#ai #decisions

[Serge Semenov @kind_serge]

Minimizing risk in predictions/planning is not always the only strategy. It's about trade-offs and objectives. High risk can yield high rewards.
#decisions #ai

[Serge Semenov @kind_serge]

Anthropic and OpenAI will not replace white-collar jobs. Have you seen any product from them that makes real-world decisions? Thus the slop.

[Serge Semenov @kind_serge]

@AnatoliKopadze This is AI slop.
I worker at Meta, and that's NOT how make decisions for serious high-stake programs.

[Serge Semenov @kind_serge]

This. Is. Deep.

[Ashwin Gopinath @ashwingop]

https://t.co/yEiQYYff21