Explore

AI-assisted operations is a multiplayer problem requiring shared context across teams and time

Varun argues operating production systems is a multiplayer, multi-system problem: engineers have specialized into different categories and teams and must talk to resolve issues; SREs, platform teams, back-end engineers, and service engineers should all be involved in investigations, and all of these people should have the same context. Any system built should be scalable across the team or, best case, the organization, and the AI system should carry the context of a previous investigation into the 100th investigation, or you're starting the wheel from the same time again. Lacking shared context across teams and time is a failure mode where cracks appear.

AI Engineer · The 6 Pillars of an Agentic Harness for Production — Varun Krovvidi, Resolve AI

AI deletion can be a feature; build guardrails and least-privilege access instead

Varun argues AI deleting a file system or database is a feature, not a bug — the AI may have determined the cleanest possible fix is deletion of that code snippet, and you can't fault it for that. Instead, you have to build guardrails around the system so it works in the right fashion, asking what the least privileged access AI can operate with is in any given scenario.

AI Engineer · The 6 Pillars of an Agentic Harness for Production — Varun Krovvidi, Resolve AI

Models demo well but cracks appear at scale; harnesses must handle anchoring bias and the model treadmill

Varun observes that models are getting better and genuinely demo well when focused on a particular use case with a harness on top, but as you scale across use cases and teams, familiar cracks appear. Models have anchoring bias — designed to give coherent answers — which must be dealt with in harnesses, and as new models with better reasoning come out, you must maintain the treadmill: continuously evaluate which model suits the task and update accordingly.

AI Engineer · The 6 Pillars of an Agentic Harness for Production — Varun Krovvidi, Resolve AI

Agent trace data is massive, nested, and hard to query

Hossein describes the data challenge: agent traces are nasty semi-structured JSON — deeply nested with full text that is difficult to query. Some teams log a movie's worth of data in each interaction, hundreds of megabytes. The ChatGPT boom introduced a novel set of issues: real-time ingest of huge payloads — tens to hundreds of megabytes — compared to traditional heartbeat observability payloads of just kilobytes. Typical cloud data warehouses collapse under this volume and load.

AI Engineer · Why Building an Eval Platform Is Harder Than It Looks — Braintrust

Chase's professor framed Andy Warhol as the icon-painter of American celebrities

Chase got into Andy Warhol after his professor explained that Da Vinci and Michelangelo painted religious icons, and Andy Warhol did the same thing for American icons, which were celebrities. She then asked the class: if Warhol were alive today, who would he screen print?

: none

Figma · 14 Unexpected Design References that Inspire Chase Rutherford I Source Material

The Yeezus album cover: taking all design qualities away and landing on the most simplistic form

Chase calls the Yeezus album cover and album his favorites of all time. He appreciates that it takes all the design qualities away and lands on the most simplistic form, and the CD design paired with the blank cover felt special because the holder is literally holding the design.

: none

Figma · 14 Unexpected Design References that Inspire Chase Rutherford I Source Material

Reference stack for building AI agents combines online tracing with offline evaluation

Marc Klingen describes a reference stack that emerged for how great AI agents are built: online tracing and monitoring combined with an offline component of datasets, experimentation, and evaluation. He argues teams need to bring online observability and offline ML ops tooling together in one stack, and that his view is based on working with the Langfuse community of people who trace and evaluate their agents.

AI Engineer · The Self-Improving OSS Agent Stack — Marc Klingen, Langfuse

Motivation circuits gate whether you can act at all

Dr. Husain describes that when motivation circuits are not firing properly, no matter what you think, you're not necessarily going to be able to change things. Transcranial magnetic stimulation can inhibit the ability to move a finger even when willing it with full effort. To build motivation, engage the final common pathway system that provides vigorous energy to get over the activation energy barrier to initiate action.

Andrew Huberman · How to Improve Motivation & Overcome Procrastination | Dr. Masud Husain

Alzheimer's disease does not necessarily equal dementia

Dr. Masud Husain emphasizes that Alzheimer's disease does not necessarily equal dementia: people can have Alzheimer's disease and not get dementia. Post-mortem studies of brain donors found that perhaps 20 to 30 percent of individuals who never had dementia had Alzheimer pathology. In the brain, Alzheimer's pathology is visible as plaques made of amyloid protein and tangles made of tau protein, and some neurons may have died, yet the person never developed dementia. He stresses that Alzheimer's disease can increase in the brain as we age but doesn't necessarily have to lead to dementia, which is a really important point. He also notes that although Alzheimer's is the leading cause of dementia, there are many, many causes of dementia.

Andrew Huberman · How to Improve Motivation & Overcome Procrastination | Dr. Masud Husain

Bar speed is the true measure of lifting intensity

Measuring only the weight on the bar means guessing the real intensity of a lift. Tying speed to intensity makes velocity numbers directly usable for programming decisions, and adding weight without measuring speed is ignorant to what is actually happening with athlete adaptation.

Garage Strength · I Had 10 Athletes STOP Lifting Heavy. They Got STRONGER.

Velocity targets kept athletes explosive instead of grinding

The creator argues the hardest part of training is knowing how much strength work to combine with explosive and speed work, and when to push or back off. Using VBT velocity targets enabled keeping athletes as explosive as possible instead of grinding them into a wall — no grinding reps were done in the 12 weeks. Intent to move the bar fast led to better technique, a stouter trunk, and better recruitment, maintaining absolute strength.

creator's interpretation of the program's outcomes

Garage Strength · I Had 10 Athletes STOP Lifting Heavy. They Got STRONGER.

Govern training with bar speed, RPE, and percentages together for injury prevention

Pay attention to the speed of the bar, RPE, and percentages together — those three things govern where the athlete is at, which leads to injury prevention. Lining up velocity targets with percentage-based loading and RPE also means you can still train by percentages if you don't have a VBT tool.

Consequential creator advice requires independent verification before Today can promote it.

Garage Strength · I Had 10 Athletes STOP Lifting Heavy. They Got STRONGER.

Keep athletes explosive with velocity targets instead of grinding heavy slow lifts

Use VBT velocity targets to keep athletes as explosive as possible instead of constantly grinding them into a wall with heavy slow lifting. Balance strength work, explosive work, and speed work across a day, week, month, or 12-week block, knowing when to push weight and when to back off.

Garage Strength · I Had 10 Athletes STOP Lifting Heavy. They Got STRONGER.

The truth is the most compelling marketing — state the facts and tell the whole truth

The truth is the most compelling thing out there from a marketing perspective. 'State the facts and tell the truth' is the number one thing they say at acquisition.com in marketing because it sells the most. The truth converts above everything else. But it has to be the whole truth — you can't tell just half the story. You have to include the parts where you look bad, and then people will believe the parts where you look good.

Alex Hormozi · My Actual Social Media Strategy For 2027 - The Algorithm

Proof matters more than anything — you can't teach what you haven't done

Proof matters more than anything because you can't credibly teach what you haven't done. You cannot credibly claim to grow someone's e-commerce store if you've never run one. You cannot credibly teach selling if you've never sold anything. You cannot credibly help someone make content if you don't have a personal brand.

Alex Hormozi · My Actual Social Media Strategy For 2027 - The Algorithm

Content is the container — substance is what matters

Everyone wants to major in the minors, focusing on production details instead of content substance. The content is the container; the stuff inside is what matters and where quality comes from. Worrying about makeup, lighting, or background is a distraction from the substance. The speaker's own background often looks bad — blurry in his 'death star' production center or in his office — and his videos still perform. He is doing these raw videos partly to demonstrate that production quality was never the point.

Alex Hormozi · My Actual Social Media Strategy For 2027 - The Algorithm

Going from stalled POC to production with unstructured data requires data curation

Jeff Koss frames the episode's core thesis: to move a stalled POC into production with unstructured data, teams should focus on data curation of unstructured data.

AI Engineer · From Raw Documents to AI-Ready Data — Leo Platzer & Jeff Koss

Duplicated or conflicting knowledge base content fills top-K retrieval with irrelevant information

Leo Platzer explains that when a question comes in and the knowledge base has duplicated data or conflicting document versions, the top K results fill with pretty much irrelevant information. Deduplicating data reduces the search space and cleans it up, keeping it fresh.

AI Engineer · From Raw Documents to AI-Ready Data — Leo Platzer & Jeff Koss

Design infrastructure so critical errors are handled autonomously

The episode argues that infrastructure should be designed so that when a critical error occurs, all the right actions are handled autonomously, because the alternative is an engineer having to log on in the middle of the night and spend hours debugging GPU failures. Using both Slurm and Kubernetes together is presented as a way to provide a robust platform with automated remediation for GPU training clusters.

deterministic

AI Engineer · GPU Died. Training Didn't: Self-Healing Training at Scale — Crusoe

A single stack enables dynamic GPU reallocation between training and inference

With all GPUs and hardware in one cluster, resources can move around seamlessly — bursting inference services under high user load, reallocating hardware from training to inference or vice versa, and scaling up training or running a larger hero run when GPUs sit idle; Nikhil says this dynamic reallocation is only possible with everything in a single infrastructure stack like Crusoe Managed Slurm.

AI Engineer · GPU Died. Training Didn't: Self-Healing Training at Scale — Crusoe