Monday, June 1, 2026
Efficiency is the new state of the art
June 1 · 8 videos
Jeff Dean predicts a 1,000,000x compute leap.
Inference now dominates 90 percent of data center workloads.
Bertrand Charpentier says efficiency is a primary dimension of performance.
Ethan He built Grok Imagine in three months.
Joe Reeve calls it vibe coding.
Startups are moving from selling tools to selling outcomes.
The network is becoming the security sandbox.
“Efficiency is a dimension of state of the art, not a footnote.”
20 days of compute vs 7 hours: rethinking what state-of-the-art means — Bertrand Charpentier, Pruna
Bertrand Charpentier · AI Engineer · 19 min
Watch on YouTube →Bertrand Charpentier challenges the focus on leaderboard rankings. He argues that efficiency must be treated as a core performance metric rather than a secondary concern.
- Human evaluation is subjective and biased by small sample sizes: manual inspection must be scaled to be reliable.
- There is no single best model because different models excel at specific tasks like object removal or background editing.
- The lazy solution of picking the top model on a leaderboard can result in 20x higher costs and latency.
- Statistical significance is often lacking in public leaderboards compared to models with millions of daily inferences.
- Efficiency is about the feasibility of scaling evaluation and deployment for complex products.
- A standard evaluation of 26,000 battles takes 20 days and $5,000 of compute for foundation models.
- Compressed models can perform the same 26,000 battle evaluation in 7 hours for $265.
How Startups Close Million Dollar Deals
Dalton Caldwell · Dalton + Michael · 18 min
Watch on YouTube →Dalton Caldwell and Michael Seibel explain why startups must target enterprise deals. They advocate for selling business outcomes rather than developer tools.
- Corporate CEOs are incentivized by revenue and stock price rather than the elegance of developer tools.
- The slowness of enterprise sales cycles allows founders to parallel-process multiple deals during gaps.
- Consumer giants like Facebook rely on 8-figure enterprise advertising deals despite their public-facing brand.
- Startups should research how target customers purchase from existing vendors to match their procurement patterns.
- Selling software as a tool becomes a race to the bottom as AI drives the cost of code toward zero.
- Selling outcomes allows startups to capture massive margin expansion as internal costs drop.
- AWS needed to hire 5,000 salespeople in 2018 to stay competitive in the enterprise market.
Inside xAI: Building Grok Imagine in 3 Months, Videogen vs World Models, and Video Agents— Ethan He
Ethan He · Latent Space · 104 min
Watch on YouTube →Ethan He discusses the rapid development of Grok Imagine at xAI. He explains why language models are the true source of intelligence in video generation.
- A small team at xAI built Grok Imagine from scratch in three months by prioritizing iteration speed.
- Visual intelligence in video models stems largely from language models acting as thinking layers.
- World models must be real-time, interactive, and long-horizon to be truly useful.
- Generative UI may replace traditional software interfaces by moving directly from user intent to pixels.
- Storing and moving petabytes of training data can cost millions in storage and egress fees.
- The current bottleneck for video generation is reasoning and agentic harnesses rather than pixel quality.
- Iterative cycle time for experiments is a better predictor of success than total compute availability.
What Happens After A 1,000,000x AI Compute Leap? | Jeff Dean
Jeff Dean · Two Minute Papers · 28 min
Watch on YouTube →Jeff Dean outlines the future of AI compute and hardware. He identifies inference and continual learning as the next major frontiers.
- Continual learning remains a major unsolved frontier where models learn from every interaction.
- The next decade of AI will be defined by multi-agent workflows solving complex engineering tasks.
- Hardware specialization for inference is now more critical than training optimization.
- Inference now accounts for roughly 90 percent of machine learning workloads in data centers.
- The data wall is likely a myth that can be overcome by algorithmic improvements and synthetic data.
- A 1,000,000x leap in compute over the next decade could enable designing an airplane in five days.
- Regulated industries like healthcare take longer to transform due to safety and privacy hurdles.
What if the network was the sandbox? — Remy Guercio, Tailscale
Remy Guercio · AI Engineer · 24 min
Watch on YouTube →Remy Guercio proposes using network identity to secure AI agents. He demonstrates how to sandbox agents without exposing sensitive credentials.
- Traditional sandboxing fails when high-privilege API keys are placed inside the isolated environment.
- Moving authentication to the network layer provides agents with placeholder keys that have no value outside the context.
- Aperture acts as an LLM gateway on a tailnet to allow granular, identity-based access control.
- Internal data shows that bash commands currently dominate over structured tool calls in agent usage.
- Centralized AI gateways allow for unified budget management across multiple providers like OpenAI and Anthropic.
- Visibility into agent behavior is often more critical for enterprise adoption than blocking specific actions.
- A simple hello request in Claude Code can cost 20 cents due to high initial context overhead.
I've Lived More In My Head Than I've Lived In Real Life
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Rob Dial explores the cost of mental absence. He provides techniques for shifting from psychological time to physical presence.
- Humans spend 47 percent of their waking lives thinking about something other than their current activity.
- Anxiety lives exclusively in the mind while peace is accessed through the physical body.
- Awareness of the moment of departure is more important than achieving perfect focus.
- Overthinking is often an intellectual disguise used to avoid the discomfort of repressed feelings.
- Productivity obsession can lead to mental absence that degrades the quality of leadership decisions.
- Arrival rituals between tasks prevent mental residue from one environment leaking into the next.
- Intentional non-narration helps break the habit of mentally rehearsing the future or replaying the past.
How to talk to statues — Joe Reeve, ElevenLabs
Joe Reeve · AI Engineer · 33 min
Watch on YouTube →Joe Reeve describes building a viral AI app in two hours. He explores the rise of vibe coding and the future of voice interfaces.
- The Statue App was built in two hours using Cursor and a single one-shot prompt.
- Vibe coding allows people without technical backgrounds to create creative software by stitching APIs.
- The story behind how a project is built can be more viral than the product itself.
- Human-to-agent interactions suffer because users are often too polite to interrupt AI agents.
- Voice UI should be paired with high-density visual feedback to overcome audio-only limitations.
- The app jumped to 1.5 million impressions after being framed as vibe coding on social media.
- Total latency from taking a photo to starting a voice call with a statue was reduced to 30 seconds.
Peptides: The Science, Uses & Safety | Dr. Abud Bakri
Abud Bakri · Andrew Huberman · 168 min
Watch on YouTube →Dr. Abud Bakri and Andrew Huberman discuss the science and risks of peptides. They cover tissue repair, metabolic health, and the unregulated gray market.
- Peptides act as a biological language mediating communication between cells as epigenetic modifiers.
- BPC-157 shows significant tissue regeneration in animal models but lacks robust human clinical data.
- The gray market for research-only compounds carries high risks of inconsistent quality or wrong substances.
- The Celebrity Trinity Stack combines GLP-1 agonists, growth hormone secretagogues, and androgen modulation.
- GLP-1s act as a powerful signal to the brain to stop eating by integrated hormonal calculations.
- Peptide patents are fragile because small changes to amino acid sequences create legally new compounds.
- Ozempic prices vary wildly, costing $1500 in the US compared to $150 in Mexico.
References
PeopleBertrand Charpentier · Dalton Caldwell (x.com/daltonc) · Michael Seibel (x.com/mwseibel) · Andy Jassy · Ethan He (x.com/EthanHe_42) · Elon Musk · Kaiming He · Jeff Dean · Jensen Huang · Demis Hassabis · Steven Balaban · Károly Zsolnai-Fehér · Remy Guercio · Rob Dial (http://coachwithrob.com) · Ram Dass · Joe Reeve (x.com/isnit0) · Sir Michael Caine · Abud Bakri (x.com/AbudBakri) · Vladimir Khavinson · Lauren Pickart · Hans Selye
ToolsPruna AI · Design Arena · Artificial Analysis · Grok Imagine · TPU · Aperture · Tailscale · WireGuard · Claude Code · Cursor · ElevenLabs Voice Design API · Ozempic · BPC-157 · GLP-1