Since Claude 4 Opus things have been refreshingly quiet. Video break!

The First Good AI Videos

First up we have Prompt Theory, made with Veo 3, which I am considering the first legitimately good AI-generated video I’ve seen. It perfectly combining form and function. Makes you think.

Here's a variant, to up the stakes a bit, then here is him doing that again.

What does it say about the medium, or about us, that these are the first legit videos?

This was the second clearly good product. Once again, we see a new form of storytelling emerging, a way to make the most of a series of clips that last a maximum of eight seconds each. The script and execution are fantastic.

I predict that will be the key for making AI videos at the current tech level. You have to have a [...]

---

Outline:

(00:16) The First Good AI Videos

(01:41) We Can Talk

(02:05) Our Price Cheap

(04:27) Changing the Game

(07:53) Making Things Interesting

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Previously: #1, #2, #3, #4.

Since we all know that dating apps are terrible, the wise person seeks to meet prospective dates in other ways, ideally in the physical world.

Alas, this has gotten more difficult. Dating apps and shifting norms mean it is considered less appropriate, and riskier, to approach strangers, especially with romantic intent, or to even ask people you know out on a date, which has a fat tail of life changing positive consequences.

People especially men are increasingly more afraid of rejection and other negative consequences, including a potential long tail of large negative consequences. Also people's skills at doing this aren’t developing, which both decreases chances of success and increases risk. So a lot of this edition is about tackling those basic questions, especially risk, rejection and fear.

There's also the question of how to be more hot and know [...]

---

Outline:

(01:20) You're Single Because Hello Human Resources

(03:02) You're Single Because You Don't Meet Anyone's Standards

(04:23) You're Single Because You Don't Know How to Open

(10:46) You're Single Because You Never Open

(15:15) You're Single Because You Don't Know How to Flirt

(21:03) You're Single Because You Won't Wear the Fucking Hat

(23:40) You're Single Because You Don't Focus On The People You Want

(29:10) You're Single Because You Choose the Wrong Hobbies

(32:54) You're Single Because You Friend Zone People

(33:44) You're Single Because You Won't Go the Extra Mile

(35:54) You're Single Because You're Overly Afraid of Highly Unlikely Consequences

(39:01) You're Single Because You're Too Afraid of Rejection

(45:49) You're Single Because You're Paralyzed by Fear

(46:37) You're Single Cause You're Not Hot Enough

(48:00) You're Single Because You Can't Tell How Hot You Look

(50:57) You're Single Because You Have the Wrong Hairstyle

(51:42) You're Single Because You're In the Wrong Place

(52:47) You're Single Because You Didn't Hire a Matchmaker

(54:13) You're Single So Here's the Lighter Side

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

How good are Claude Opus 4 and Claude Sonnet 4?

They’re good models, sir.

If you don’t care about price or speed, Opus is probably the best model available today.

If you do care somewhat, Sonnet 4 is probably best in its class for many purposes, and deserves the 4 label because of its agentic aspects but isn’t a big leap over 3.7 for other purposes. I have been using 90%+ Opus so I can’t speak to this directly. There are some signs of some amount of ‘small model smell’ where Sonnet 4 has focused on common cases at the expense of rarer ones. That's what Opus is for.

That's all as of when I hit post. Things do escalate quickly these days, although I would not include Grok in this loop until proven otherwise, it's a three horse race and if you told me [...]

---

Outline:

(01:17) On Your Marks

(05:32) Standard Silly Benchmarks

(11:09) API Upgrades

(12:45) Coding Time Horizon

(13:47) The Key Missing Feature is Memory

(14:52) Early Reactions

(26:12) Opus 4 Has the Opus Nature

(32:27) Unprompted Attention

(35:09) Max Subscription

(36:24) In Summary

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Unlike everyone else, Anthropic actually Does (Some of) the Research. That means they report all the insane behaviors you can potentially get their models to do, what causes those behaviors, how they addressed this and what we can learn. It is a treasure trove. And then they react reasonably, in this case imposing their ASL-3 safeguards on Opus 4. That's right, Opus. We are so back.

Yes, there are some rather troubling behaviors that Opus can do if given the proper provocations. If you tell it to ‘take initiative,’ hook it up to various tools, and then tell it to fabricate the data for a pharmaceutical study or build a bioweapon or what not, or fooling Opus into thinking that's what you are doing, it might alert the authorities or try to cut off your access. And That's Terrible, completely not intended behavior, we agree it shouldn’t do [...]

---

Outline:

(03:44) Introducing Claude 4 Opus and Claude 4 Sonnet

(06:08) Activate Safety Level Three

(08:17) The Spirit of the RSP

(09:46) An Abundance of Caution

(10:34) Okay What Are These ASL-3 Precautions

(16:55) How Annoying Will This ASL-3 Business Be In Practice?

(18:34) Overview Of The Safety Testing Process

(20:12) False Negatives On Single-Turn Requests

(20:39) False Positives on Single-Turn Requests

(21:33) Ambiguous Requests and Multi-Turn Testing.

(22:21) Child Safety

(22:55) Political Sycophancy and Discrimination

(25:46) Agentic Safety Against Misuse

(26:54) Alignment

(27:40) The Clearly Good News

(30:40) Reasoning Faithfulness Remains Unchanged

(31:47) Self-Preservation Attempts

(44:12) High Agency Behavior

(01:10:37) Oh Now You Demand Labs Take Responsibility For Their Models

(01:18:32) In The Beginning The Universe Was Created, This Made a Lot Of People Very Angry And Has Been Widely Regarded as a Bad Move

(01:30:52) Erratic Behavior and Stated Goals in Testing

(01:36:43) Situational Awareness

(01:37:11) Insufficiently Mostly Harmless Due To Then-Omitted Data

(01:40:15) Apollo Evaluation

(01:43:36) Model Welfare

(01:48:50) The RSP Evaluations and ASL Classifications

(01:55:22) Pobody's Nerfect

(01:56:54) Danger, And That's Good Actually

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

What a week, huh? America signed a truly gigantic chip sales agreement with UAE and KSA that could be anything from reasonable to civilizational suicide depending on security arrangements and implementation details, Google announced all the things, OpenAI dropped Codex and also bought Jony Ive's device company for $6.5 billion, Vance talked about reading AI 2027 (surprise, in a good way!) and all that other stuff.

Lemon, it's Thursday, you’ve got movie tickets for Mission Impossible: Final Reckoning (19th and Broadway AMC, 3pm), an evening concert tonight from Light Sweet Crude and there's a livestream from Anthropic coming up at 12:30pm eastern, the non-AI links are piling up and LessOnline is coming in a few weeks. Can’t go backwards and there's no time to spin anything else out of the weekly. Got to go forward to go back. Better press on.

So for the moment, here we [...]

---

Outline:

(01:11) Language Models Offer Mundane Utility

(04:43) Language Models Don't Offer Mundane Utility

(09:12) Huh, Upgrades

(09:30) Codex of Ultimate Vibing

(11:37) On Your Marks

(11:55) Choose Your Fighter

(14:01) Deepfaketown and Botpocalypse Soon

(18:33) Copyright Confrontation

(19:34) Regarding South Africa

(20:01) Cheaters Gonna Cheat Cheat Cheat Cheat Cheat

(25:00) They Took Our Jobs

(32:47) The Art of the Jailbreak

(32:54) Get Involved

(33:50) In Other AI News

(37:45) Much Ado About Malaysia

(46:21) Show Me the Money

(47:23) Quiet Speculations

(57:00) Autonomous Dancing Robots

(01:00:32) The Quest for Sane Regulations

(01:05:29) The Mask Comes Off

(01:11:28) The Week in Audio

(01:13:39) Write That Essay

(01:14:37) Vance on AI

(01:22:57) Rhetorical Innovation

(01:31:10) Margaritaville

(01:35:53) Rhetorical Lack of Innovation

(01:38:19) If Anyone Builds It, Everyone Dies

(01:50:14) Aligning a Smarter Than Human Intelligence is Difficult

(01:53:30) People Are Worried About AI Killing Everyone

(01:55:58) The Lighter Side

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

What did Google announce on I/O day? Quite a lot of things. Many of them were genuinely impressive. Google is secretly killing it on the actual technology front.

Logan Kilpatrick (DeepMind): Google's progress in AI since last year:

– The worlds strongest models, on pareto frontier

– Gemini app: has over 400M monthly active users

– We now process 480T tokens a month, up 50x YoY

– Over 7M developers have built with the Gemini API (4x)

Much more to come still!

I think? It's so hard to keep track. There's really a lot going on right now, not that most people would have any idea. Instead of being able to deal with all these exciting things, I’m scrambling to get to it all at once.

Google AI: We covered a LOT of ground today. Fortunately, our friends at @NotebookLM put [...]

---

Outline:

(03:21) The TLDR

(05:18) Flow, Veo 3 and Imagen 4

(08:39) Gmail Integration That's Actually Good?

(11:41) Gemini 2.5 Flash

(15:08) Gemma 3n

(16:59) Gemini Diffusion

(17:53) Jules

(18:36) We're in Deep Research

(19:28) Google Search 'AI Mode'

(22:19) AI Shopping

(23:05) Agent Mode

(24:10) Project Astra or is it Google Live?

(25:26) Android XR Glasses

(26:16) Gemini For Your Open Tabs In Chrome

(26:48) Google Meet Automatic Translation

(27:53) We Have Real 3D At Home, Oh No

(30:10) You Will Use the AI

(31:45) Our Price Cheap

(35:06) What To Make Of All This

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

While we wait for wisdom, OpenAI releases a research preview of a new software engineering agent called Codex, because they previously released a lightweight open-source coding agent in terminal called Codex CLI and if OpenAI uses non-confusing product names it violates the nonprofit charter. The promise, also reflected in a number of rival coding agents, is to graduate from vibe coding. Why not let the AI do all the work on its own, typically for 1-30 minutes?

The answer is that it's still early days, but already many report this is highly useful.

Introducing Codex

Sam Altman: today we are introducing codex.

it is a software engineering agent that runs in the cloud and does tasks for you, like writing a new feature of fixing a bug.

you can run many tasks in parallel.

it is amazing and exciting how much software one [...]

---

Outline:

(00:43) Introducing Codex

(04:24) System Card Addendum

(06:37) Update to Codex CLI

(07:50) Overall Reception

(09:36) Codex Offers Mundane Utility

(15:25) Codex Doesn't Offer Mundane Utility

(18:36) Our Price Cheap

(20:38) Two Kinds of Agent

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Our government, having withdrawn the new diffusion rules, has now announced an agreement to sell massive numbers of highly advanced AI chips to UAE and Saudi Arabia (KSA). This post analyzes that deal and that decision.

It is possible, given sufficiently strong agreement details (which are not yet public and may not be finalized) and private unvoiced considerations, that this deal contains sufficient safeguards and justifications that, absent ability to fix other American policy failures, this decision is superior to the available alternatives. Perhaps these are good deals, with sufficiently strong security arrangements that will actually stick.

Perhaps UAE and KSA are more important markets and general partners than we realize, and the rest of the world really is unable to deploy capital and electrical power the way they can and there is nothing we can do to change this, and perhaps they have other points of [...]

---

Outline:

(02:28) Choosing Sides In the War on Cancer

(03:23) The Central Points From Last Week

(07:34) Diffusion Controls Have Proven Vital

(10:27) It's a Huge Deal

(14:11) Do You Feel Secure?

(17:09) Why just count server racks fails

(18:29) Bottom-line probability estimate

(21:41) Semianalysis Defends the Deal

(28:09) Understanding the China Hawks

(35:35) Rhetoric Unbecoming

(37:18) Could China Have 'Done This Deal'?

(38:37) Tyler Cowen Asks Good Questions

(43:38) Saudi Arabia Also Made a Deal

(45:35) At Best A Second Best Solution

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

A Funny Thing Happened on Twitter

A Brief History of Similar Incidents

---

Outline:

(00:22) A Funny Thing Happened on Twitter

(05:54) A Brief History of Similar Incidents

(08:29) Speculations on What Happened

(16:23) The Jabs Continue

(17:11) Show Us Your System Prompts

(19:15) A Perfectly Reasonable Explanation For All This

(26:23) How Should We Think About This Going Forward?

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

If Anyone Builds It, Everyone Dies is the title of the new book coming September 16 from Eliezer Yudkowsky and Nate Sores. The ‘it’ in question is superintelligence built on anything like the current AI paradigm, and they very much mean this literally. I am less confident in this claim than they are, but it seems rather likely to me. If that is relevant to your interests, and it should be, please consider preordering it.

This week also featured two posts explicitly about AI policy, in the wake of the Senate hearing on AI. First, I gave a Live Look at the Senate AI Hearing, and then I responded directly to arguments about AI Diffusion rules. I totally buy that we can improve upon Biden's proposed AI diffusion rules, especially in finding something less complex and in treating some of our allies better, no one is saying we [...]

---

Outline:

(02:33) Language Models Offer Mundane Utility

(06:24) Language Models Don't Offer Mundane Utility

(06:55) Huh, Upgrades

(08:07) Gemini 2.5 Pro Gets An Ambiguous Upgrade

(15:30) GPT-4o Is Still A (Less) Absurd Sycophant

(20:17) Choose Your Fighter

(20:35) Deepfaketown and Botpocalypse Soon

(20:52) Copyright Confrontation

(26:52) Cheaters Gonna Cheat Cheat Cheat Cheat Cheat

(28:25) They Took Our Jobs

(30:16) Safety Third

(31:00) The Art of the Jailbreak

(31:09) Get Involved

(32:04) If Anyone Builds It, Everyone Dies

(34:42) Endorsements for Eliezer's Book

(37:11) Why Preorders Matter

(41:14) Great Expectations

(42:07) Introducing

(45:51) In Other AI News

(48:27) Quiet Speculations

(54:35) Four Important Charts

(56:52) Unprompted Suggestions

(01:02:11) Unprompted Suggestions For You

(01:03:06) How to Be a Good Claude

(01:04:00) The Quest for Sane Regulations

(01:14:44) The Week in Audio

(01:14:53) Rhetorical Innovation

(01:20:53) Aligning a Smarter Than Human Intelligence is Difficult

(01:21:34) Is the Pope Worried About AI

(01:22:27) People Are Worried About AI Killing Everyone

(01:23:49) The Lighter Side

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.