An important part of Geffen’s advice is the implied trial and error. Acting will allow you to experiment with a few options instead of stalling in analysis paralysis.

First PR submitted to @OmarchyLinux Packages. Updating Grok Bot (@bot) to v0.63. The current package is pinned to v0.29.

This is a common strategy. Downplay the competition while you try to build your own version of the product. Case in point: AT&T and AST.

Be careful to take this at face value. A few notes from this study to be aware of: 1. The founders were already on an outside-capital path. 2. Wasserman dropped people who refused any outside money. 3. The “~50% more” number is paper value (last-round pre-money × founder %), not cash from a sale. 4. Realized IPO/acquisition proceeds were left for later work.

The Self-Deception of Evaluation One of the hardest traps to avoid is self-deception. Often, we fall into its jaws without a fight and become a hostage who defends the kidnapper. And you can only see the cell after you're out of it. Present day AI is optimized to be evaluated by humans. It impresses its adjudicators, and that masks which tasks it should be doing and still isn’t. Goodhart's Law says that when a measure becomes a target, it stops being a good measure. If the focus is hitting a number, the behavior of the system changes in ways that compromise what the metric was supposed to measure. AI is a wonderful technology and has the potential to change how the world functions. And although a lot has been promised, much less has been delivered as products people use reliably. Processes that are done by many, but loved by few are perfect candidates to be handled by AI, yet it falls short in delivering solutions. For example: account reconciliation, project task prioritization, expense filing, fraud detection, and task assignment. My point is: Are the current metrics used to evaluate AI good enough for the job? Or are we fooling ourselves with numbers that are optimized to impress us?

A year ago this was an internship project, then a conference paper, and now it is published in Springer Nature. Pierce Buckner-Wolfson and I researched and wrote “Using Embeddings to Train and Build Privacy-Preserving Machine Learning Models and Applications.” It addresses a critical constraint most product leaders already know. You need data to build anything useful, but you cannot treat raw data as a free input. This work shows you can train practical models on embeddings alone. Search, anomaly detection, recommendations, sentiment. No raw records required. If a project of yours keeps getting stalled because access to raw data is not permitted, this is the kind of solution worth implementing, so you can keep building while maintaining compliance. doi.org/10.1007/978-3-032-28… Have you killed, delayed, or watered down a product because the data was off limits?

A coffee shop can be a good place to work for a few hours. The many sounds blend and become a sort of white noise. But sometimes there's that insufferable person who takes a call and talks loudly, annoying everyone in the place. Worse yet, they know it.

Conspiracies require too much competence. Although some are true, most are rationalization stories after a series of small unaddressed tragedies.

Once it’s honest and from someone you trust, feedback becomes a mechanism for error correction, just like the small adjustments on the steering wheel that keep you on the road.

The Hobbit would have been a much shorter story if my dog had joined Bilbo and the dwarves.

This holds true even if the argument is reversed. A lack of constraints would mean infinite time to try everything. That in itself is constrained by our lifetimes. Constraints are the chainsaws that prune the branches of infinite options.

When Selling a Vision Looks Like Snake Oil: Why Some Predictions Land, Most Miss the Mark, and How Eigenvectors Explain the Hits.

When Selling a Vision Looks Like Snake Oil. An overhyped launch and an underwhelming delivery. Why some predictions land, while most miss the mark. There is a reason. Essay goes live tomorrow.

None of the features of the iPhone Duo create focus. They’re pretty much all about multitasking and more media consumption. Ternus said the device is meant to be a hub, but a hub of what? He mentioned it, but didn’t define what a hub was.

Using the consequences as an explanation to justify the causes is a distraction technique. It only adds confusion to the discussion. Solid decision-making is built on your ability to explain the reasoning behind the decision. For instance: "We cannot find senior people." Often the consequence of never hiring juniors, never teaching the work, and running a process that only accepts people who already did the job somewhere else. "Engineering is slow." Often the consequence of unclear scope, too much work in flight, and priorities that change midway. "Customers do not use that feature." Often the consequence of burying it, shipping it broken, or never explaining it. Non-use gets treated as proof the idea was wrong. "We need this meeting or nobody stays aligned." Often the consequence of the other twenty unnecessary meetings. "Nobody on this team takes initiative." Often the consequence of decisions getting pulled back upstream every time someone tried. The litmus test for any of these is to remove the outcome and ask whether the explanation still stands. When it collapses, you're staring at a consequence.

Opinion: Shouting AGI is about investor liquidity. In its original agreement with Microsoft, OpenAI would no longer have to share the rights to new models with them after there was sufficient proof of AGI. However, that clause is no longer part of the agreement between them since April 2026. But investors and employees still seek liquidity for their investments, so shouting "AGI" when, by OpenAI's own definition, the model doesn't outperform humans at most economically valuable work, benefits those looking to cash out. The model is impressive in itself without being AGI. Those on the bullhorn need the claim more than industry or customers do.

Speak last when a decision hasn’t been made yet. Otherwise, you”re setting the tone for the discussion and potentially ending it. Speak first when a decision has been made already. Communicate clearly, then stay for the questions.

For those who prefer natural scroll direction with mouse and trackpad when using @OmarchyLinux nvim ~/.config/hypr/input.lua hl.config({ input = { natural_scroll = true, touchpad = { natural_scroll = true, }, }, }) :wq