Explore

Lessons at the student's level solve the IQ issue; then the student must do the work

Joe Liemandt argues that if you give a student a lesson at their level, they can learn it — a seventh grader who needs fourth-grade material cannot learn seventh-grade lessons but can learn the fourth-grade lesson. US Common Core is designed so that everyone can learn it — the conceptual leaps are not that big. With material delivered at the right level, you can get everybody through eighth grade, and delivering lessons at the student's level solves the IQ issue of whether they can learn the material. After level-appropriate lessons, the student still needs to actually do the work — go to math and grind through the lessons.

Andrew Huberman · How to Accelerate Learning & Improve Education | Joe Liemandt

Moving on at 80% mastery compounds problems

Liemandt argues that moving a student to the next grade level after finishing fourth grade at 80% mastery is wrong because the problems compound, and describes that in America students scoring at a B level (80%) are moved on to advanced material.

Andrew Huberman · How to Accelerate Learning & Improve Education | Joe Liemandt

The system's failure: delivering grade-level content to kids who lack the prerequisites

Liemandt diagnoses the current education system's core problem as delivering eighth graders eighth grade content, because an eighth grade teacher by law must stand up and deliver eighth grade content. In any classroom there are kids who need seventh, sixth, or fifth grade material, and because of COVID some who need fourth and third grade material due to learning loss. Some kids never learned to read in third and fourth grade and are now being told eighth grade content, which is why the curves are going down. He also blames the removal of basics like memorizing multiplication tables and phonics-based science of reading as lowered standards that kill kids in higher grades; he believes (uncertainly) that educators didn't realize how important these basics were.

partial (rationale for removing basics is uncertain)

Andrew Huberman · How to Accelerate Learning & Improve Education | Joe Liemandt

Why a holding company is the only natural structure for AI-era refounding

Michael Lee argues the only natural way to partner with world-class organizations, bring frontier engineering to bear, and build a shared platform across portfolio companies is through a holding company. A holding company aligns the duration of capital with the sustained investment and operational commitments required to drive transformation end-to-end, creates an organization culturally oriented around forging market leaders rather than deploying capital, and enables using retained earnings to support existing companies and net new investments.

creator's argument for his own model

No Priors: Artificial Intelligence | Technology | Startups · Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee

Models are designed to please and to give coherent answers, not causality — production needs a causal chain of evidence

Varun argues models are there to please you — push hard enough and they will agree with you in every direction — and are designed to give coherent answers but not causality. For production incidents, the only thing you're looking for is a causal chain of evidence, like a detective: the steps that led to the incident, so you fix the right issue instead of creating a patch. Solving production reasoning requires causal reasoning built into the system, not just coherent answers.

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

Context is a double bind: over-exploration with too much, under-exploration with too little — and telemetry is infinite

Varun argues context windows are not the only problem: with large enough context, models over-explore, generating hypotheses that don't exist or won't lead to the right answer; with very limited context, they under-explore and miss pathways. For production incidents, context becomes a huge issue because telemetry is literally infinite — you can create as many log lines and metrics as you want — and when infinite telemetry is married with code and infrastructure, real cracks in the system start to appear.

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

Every workflow has hundreds of task types; associating one model to a workflow is an overlooked mistake

Varun argues that associating a single model to a workflow or outcome is an overlooked mistake: every workflow has hundreds of different types of tasks — reasoning, deterministic, image reasoning, reasoning across images, SQL, or logs — and each frontier model is good at a specific kind of task. The orchestration engine needs to keep pace with how models evolve and marry the best model to the best task.

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

Track average velocity for absolute strength, peak velocity for technical coordination

For absolute strength movements like back squat, front squat, bench press, and deadlift, measure average (mean) velocity, which captures the speed of the entire rep including eccentric and concentric portions. For technical coordination movements like cleans, snatches, jerks, and push presses, measure peak velocity to capture the highest instantaneous bar speed.

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

Original films pay an 'education tax' in marketing that remakes skip

For an original film like TAR, marketing money must go to making people care about something they've never heard of — an 'education tax' that remakes avoid entirely.

Slidebean · Why Hollywood Keeps Making $125M Flops

POCs succeed on ~40 pre-curated files but break at scale on discovery, sensitivity, and quality

A POC built on around 40 pre-curated files can validate parsing, chunking, OCR, embedding, vector DB, and knowledge graph, but when scaling to 80,000+ files the bottleneck is file discovery, sensitivity scanning, and data quality — not the AI model itself. Plan for scale from the start.

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

Oversized decoders let silicon play the game while the cells never learn — an easy way to cheat

An oversized decoder can play the game by itself — if you zero out the spikes, the decoder can still play with enough biases. If the silicon decoder overfits, the cells never learn — they can keep their stochastic behavior while the decoder plays the game for them. It is very easy to cheat performance out of these systems: massively oversize your decoder, run backpropagation on it, and suddenly you have brain cells that play anything. Startups and companies get this wrong a lot of the time. In fly-brain connectome experiments, people use massively oversized decoders, so it is never the fly doing anything, just silicon. The fix was to massively undersize the decoder, run ablations to confirm the spikes carry information, and set the linear decoder's bias to zero; undersizing gave better

Presented as the speaker's strong claim based on their experience; the fly connectome example is attributed to 'people' generally.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Structure focus bouts around 90-minute ultradian cycles

Huberman states the ideal duration for any deep focus bout — physical exercise, skill or musical learning, therapy, classes, or deep conversation — is approximately 90 minutes, because brain and body operate in 90-minute ultradian cycles within each 24-hour day. He advises keeping each bout to 90 minutes or less and determining how many cycles you can perform per day based on sleep quality, nourishment, and how trained your focus capacity is.

Andrew Huberman · Tools to Improve Your Focus & Concentration | Huberman Lab Essentials

First sleep cycle delivers a big bolus of growth hormone

Growth hormone is released all day and all night long, but the first deep slow-wave sleep cycle produces a big bolus of growth hormone release in men and women equally. If you miss the first deep slow-wave sleep period, you also miss the big bolus of growth hormone release. According to endocrinologists, big boluses of growth hormone release do different things than a little bit eked out over time.

Andrew Huberman · Essentials: Use Sleep to Enhance Learning, Memory & Emotional State | Dr. Gina Poe

Implement a multi-layer security architecture for AI systems

Anthropic's approach involves many layers of security: Model Training, then Probes/Classifiers, then Auto Mode, then Identity/Permissions — a defense-in-depth architecture for AI systems.

endorsed

Latent.Space · Claude Code’s Next Era — Thariq Shihipar, Anthropic

When asked a predictable question, give a playful, surprising answer instead of the expected one

The creator advises that when someone asks the same news-based questions, don't give the same predictable answer — do the opposite. Give a playful, non-literal, non-serious answer that surprises the other person. The skill is lowering your filter and saying the first playful thing that comes to mind instead of the expected thing everyone else is reaching for.

Charisma on Command · Give Me 7 Minutes and I’ll Make You Dangerously Good at Small Talk

Sequence ab exercises bottom-up: legs first, torso last

Jeff argues that the sequence of ab exercises matters more than the exercises themselves. He recommends focusing on what is actually moving in each exercise: start with bottom-up movements where only the legs move, since these are the hardest because you must lift the weight of your legs and need the most energy early in the workout. Next come exercises where the legs and top both move, which can be done once slightly fatigued because the obliques provide additional help. Finish with top-only movements, which can be performed in any state of fatigue without compromising the workout's effect.

ATHLEAN-X™ · I'm 51. Here's How I Still Have Visible Abs (WORKS AT ANY AGE)

Embodiment makes the audience feel the emotional ride more intensely

Embodiment is Matthew McConaughey's most important trick for making the audience feel the emotional ride more intensely. When telling a story, go into the characters you play by adopting their accent and gesticulating to increase tension. Speak in the present tense when telling a story.

Charisma on Command · Matthew McConaughey’s #1 Charisma Skill

The false end: a story that could satisfyingly end, then escalates

The false end is when you've told a story so well it could be over and the audience would be satisfied, but it actually gets more intense. The narrator calls it one of the best and most underused storytelling tactics — he'd put it at the number two slot for things to use regularly — and notes people go wild for false endings; it can completely make a story. Analogies: in video games it's the boss fight with multiple phases; in keynote speeches it's Steve Jobs saying 'And one more thing.'

Charisma on Command · Matthew McConaughey’s #1 Charisma Skill

The model harness — domain data, feedback loops, scoring — separates a POC from a production product

Andrew Bialecki argues that the harness built around the model — domain-specific data, feedback loops, and scoring — is what turns a POC into a product that works for 200,000 customers, and that building AI applications at scale requires rock-solid performance across long-tail use cases.

The Official SaaStr Podcast: SaaS | Founders | Investors · SaaStr 873: Agents Are Your New Power Users: How Klaviyo CEO Andrew Bialecki Is Remaking a $1.4B Business for the Agent Era

Harvey's agent builder comes in natural language and deterministic types

Harvey's agent builder comes in two types: a natural language builder based on skills and datasets that the user selects, and a deterministic builder that follows exact steps for a given scenario.

The Official SaaStr Podcast: SaaS | Founders | Investors · SaaStr 876: Shipping Enterprise AI Agents with the CPOs of Rubrik, Glean, and Harvey