LessWrong posts by zvi

Anthropic first gave us Claude Opus 4.6, then followed up with Claude Sonnet 4.6.

For most purposes Sonnet 4.6 is not as capable as Opus 4.6, but it is not that far behind, it would have been fully frontier-level a few months ago, and it is faster and cheaper than Opus.

That has its advantages, including that Sonnet is in the free plan, and it seems outright superior for computer use.

Anthropic: Claude Sonnet 4.6 is available now on all plans, Cowork, Claude Code, our API, and all major cloud platforms.

We’ve also upgraded our free tier to Sonnet 4.6 by default—it now includes file creation, connectors, skills, and compaction.

Claude Sonnet 4.6 is our most capable Sonnet model yet. It's a full upgrade of the model's skills across coding, computer use, long-context reasoning, agent planning, knowledge work, and design. Sonnet 4.6 also features a 1M token context window in beta.

JB: I use it all the time because I’m poor.

This substantially upgrades Claude's free tier for coding and computer use. It gives us all a better lightweight option, including for sub-agents where you would have previously needed to use Haiku. [...]

---

Outline:

(01:53) On Your Marks

(10:09) Reactions: Its How Much You Save

(18:56) Bringing It Together

---

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.

This was the week of Claude Opus 4.6, and also of ChatGPT-5.3-Codex. Both leading models got substantial upgrades, although OpenAI's is confined to Codex. Once again, the frontier of AI got more advanced, especially for agentic coding but also for everything else.

I spent the week so far covering Opus, with two posts devoted to the extensive model card, and then one giving benchmarks, reactions, capabilities and a synthesis, which functions as the central review.

We also got GLM-5, Seedance 2.0, Claude fast mode, an app for Codex and much more.

Claude fast mode means you can pay a premium to get faster replies from Opus 4.6. It's very much not cheap, but it can be worth every penny. More on that in the next agentic coding update.

One of the most frustrating things about AI is the constant goalpost moving, both in terms of capability and safety. People say ‘oh [X] would be a huge deal but is a crazy sci-fi concept’ or ‘[Y] will never happen’ or ‘surely we would not be so stupid as to [Z]’ and then [X], [Y] and [Z] all happen and everyone shrugs as if nothing happened and [...]

---

Outline:

(02:32) Language Models Offer Mundane Utility

(03:17) Language Models Dont Offer Mundane Utility

(03:33) Huh, Upgrades

(04:22) On Your Marks

(06:23) Overcoming Bias

(07:20) Choose Your Fighter

(08:44) Get My Agent On The Line

(12:03) AI Conversations Are Not Privileged

(12:54) Fun With Media Generation

(13:59) The Superb Owl

(22:07) A Word From The Torment Nexus

(26:33) They Took Our Jobs

(35:36) The Art of the Jailbreak

(35:48) Introducing

(37:28) In Other AI News

(42:01) Show Me the Money

(43:05) Bubble, Bubble, Toil and Trouble

(53:38) Future Shock

(56:06) Memory Lane

(57:09) Keep The Mask On Or Youre Fired

(58:35) Quiet Speculations

(01:03:42) The Quest for Sane Regulations

(01:06:09) Chip City

(01:09:46) The Week in Audio

(01:10:06) Constitutional Conversation

(01:11:00) Rhetorical Innovation

(01:19:26) Working On It Anyway

(01:22:17) The Thin Red Line

(01:23:35) Aligning a Smarter Than Human Intelligence is Difficult

(01:30:42) People Will Hand Over Power To The AIs

(01:31:50) People Are Worried About AI Killing Everyone

(01:32:40) Famous Last Words

(01:40:15) Other People Are Not As Worried About AI Killing Everyone

(01:42:41) 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.

Things that are being pushed into the future right now:

We are officially a bit behind and will have to catch up next week.

Even without all that, we have a second highly full plate today.

Table of Contents

(As a reminder: bold are my top picks, italics means highly skippable)

---

Outline:

(00:43) Levels of Friction

(04:55) The Art Of The Jailbreak

(06:16) The Quest for Sane Regulations

(12:09) People Really Hate AI

(18:22) A Very Bad Paper

(25:21) Rhetorical Innovation

(32:35) The Most Forbidden Technique

(34:10) Everyone Is Or Should Be Confused About Morality

(36:07) Aligning a Smarter Than Human Intelligence is Difficult

(44:51) Well Just Call It Something Else

(47:18) Vulnerable World Hypothesis

(51:37) Autonomous Killer Robots

(53:18) People Will Hand Over Power To The AIs

(57:04) People Are Worried About AI Killing Everyone

(59:29) Other People Are Not Worried About AI Killing Everyone

(01:00:56) 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.

There was way too much going on this week to not split, so here we are. This first half contains all the usual first-half items, with a focus on projections of jobs and economic impacts and also timelines to the world being transformed with the associated risks of everyone dying.

Quite a lot of Number Go Up, including Number Go Up A Lot Really Fast.

Among the thing that this does not cover, that were important this week, we have the release of Claude Sonnet 4.6 (which is a big step over 4.5 at least for coding, but is clearly still behind Opus), Gemini DeepThink V2 (so I could have time to review the safety info), release of the inevitable Grok 4.20 (it's not what you think), as well as much rhetoric on several fronts and some new papers. Coverage of Claude Code and Cowork, OpenAI's Codex and other things AI agents continues to be a distinct series, which I’ll continue when I have an open slot.

Most important was the unfortunate dispute between the Pentagon and Anthropic. The Pentagon's official position is they want sign-off from Anthropic and other AI companies on ‘all legal uses’ [...]

---

Outline:

(02:26) Language Models Offer Mundane Utility

(02:49) Language Models Dont Offer Mundane Utility

(06:11) Terms of Service

(06:54) On Your Marks

(07:50) Choose Your Fighter

(09:19) Fun With Media Generation

(12:29) Lyria

(14:13) Superb Owl

(14:54) A Young Ladys Illustrated Primer

(15:03) Deepfaketown And Botpocalypse Soon

(17:49) You Drive Me Crazy

(18:04) Open Weight Models Are Unsafe And Nothing Can Fix This

(21:19) They Took Our Jobs

(26:53) They Kept Our Agents

(27:42) The First Thing We Let AI Do

(37:47) Legally Claude

(40:24) Predictions Are Hard, Especially About The Future, But Not Impossible

(46:08) Many Worlds

(48:45) Bubble, Bubble, Toil and Trouble

(49:31) A Bold Prediction

(49:55) Brave New World

(53:09) Augmented Reality

(55:21) Quickly, Theres No Time

(58:29) If Anyone Builds It, We Can Avoid Building The Other It And Not Die

(01:00:18) In Other AI News

(01:04:03) Introducing

(01:04:31) Get Involved

(01:07:15) Show Me the Money

(01:08:26) The Week In Audio

---

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.

There really is a lot going on these days.

I held off posting this because I was trying to see if I could write a net helpful post about the current situation involving Anthropic and the Pentagon. Anthropic very much wants to help DoW defend our country and make us strong. It is clear there have been some large misunderstandings here about how LLMs work.

They are not ordinary tools like spreadsheets that automatically do whatever the user asks, nor would it be safe to make them so, nor do they predictably adhere to written rule sets or take instructions from their CEO in a crisis. And they are probabilistic. You do not and cannot get absolute guarantees.

The only way to know if an AI model will do what you need in a crisis is something you needed to be do regardless of potential refusals, and which is also what you must do with human soldiers, which is to run the simulations and mock battles and drills and tests that tell you if the model can do and is willing to do the job.

If there are irreconcilable differences and the military contract needs [...]

---

Outline:

(02:09) Bad News

(04:03) Government Working

(15:11) The Epstein Files

(17:17) RIP Scott Adams

(19:08) News You Cant Use But Click On Anyway

(20:22) Were Putting Together A Team

(20:47) You Cant Retire, I Quit

(23:41) Jones Act Watch

(33:24) Variously Effective Altruism

(34:41) They Took Our Jobs And Now I Can Relax

(35:39) While I Cannot Condone This

(37:32) Good News, Everyone

(39:19) Use Your One Time

(42:34) Hands Off My Phone

(45:14) Fun Theory

(46:46) Good Advice

(48:31) For Your Entertainment

(51:07) Plur1bus

(52:14) Gamers Gonna Game Game Game Game Game

(57:58) Sports Go Sports

(01:02:49) The Revolution of Retroactive Rising Expectations

(01:04:15) I Was Promised Spying Cars

(01:05:13) Prediction Market Madness

(01:09:02) 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.

Some podcasts are self-recommending on the ‘yep, I’m going to be breaking this one down’ level. This was one of those. So here we go.

As usual for podcast posts, the baseline bullet points describe key points made, and then the nested statements are my commentary. Some points are dropped.

If I am quoting directly I use quote marks, otherwise assume paraphrases.

Normally I keep everything to numbered lists, but in several cases here it was more of a ‘he didn’t just say what I think he did did he’ and I needed extensive quotes.

In addition to the podcast, there were some discussions around safety, or the lack thereof, at xAI, and Elon Musk went on what one can only describe as megatilt, including going hard after Anthropic's Amanda Askell. I will include that as a postscript.

I will not include recent developments regarding Twitter, since that didn’t come up in the interview.

I lead with a discussion of bounded distrust and how to epistemically consider Elon Musk, since that will be important throughout including in the postscript.

What are the key takeaways?

---

Outline:

(02:56) Bounded Distrust

(05:12) IN SPACE

(09:56) The AI Will Follow You To Mars

(22:32) xAI Business Plans

(25:54) Optimus Prime

(27:04) Beating China

(30:02) SpaceX and How To Run a Company Elon Style

(33:17) DOGE

(35:29) TeraFab IN SPACE

(35:47) Postscript: Safety Third at xAI

(40:15) Elon Serves Back Saying That Which Is Not

(42:51) Elons Army

(43:55) Children Are Our Future

(48:11) Where Do We Go From Here

---

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.

Some podcasts are self-recommending on the ‘yep, I’m going to be breaking this one down’ level. This was very clearly one of those. So here we go.

As usual for podcast posts, the baseline bullet points describe key points made, and then the nested statements are my commentary. Some points are dropped.

If I am quoting directly I use quote marks, otherwise assume paraphrases.

What are the main takeaways?

---

Outline:

(01:47) The Pace of Progress

(08:56) Continual Learning

(13:46) Does Not Compute

(15:29) Step Two

(22:58) The Quest For Sane Regulations

(26:08) Beating China

---

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.

OpenAI is back with a new Codex model, released the same day as Claude Opus 4.6.

The headline pitch is it combines the coding skills of GPT-5.2-Codex with the general knowledge and skills of other models, along with extra speed and improvements in the Codex harness, so that it can now handle your full stack agentic needs.

We also got the Codex app for Mac, which is getting positive reactions, and quickly picked up a million downloads.

CPT-5.3-Codex is only available inside Codex. It is not in the API.

As usual, Anthropic's release was understated, basically a ‘here's Opus 4.6, a 212-page system card and a lot of benchmarks, it's a good model, sir, so have fun.’ Whereas OpenAI gave us a lot less words and a lot less benchmarks, while claiming their model was definitely the best.

OpenAI: GPT-5.3-Codex is the most capable agentic coding model to date, combining the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2. This enables it to take on long-running tasks that involve research, tool use, and complex execution.

Much like a colleague, you can steer and interact with GPT-5.3-Codex while [...]

---

Outline:

(01:50) The Overall Picture

(03:00) Quickly, Theres No Time

(04:15) System Card

(04:49) AI Box Experiment

(05:22) Maybe Cool It With Rm

(07:02) Preparedness Framework

(11:14) Glass Houses

(12:16) OpenAI Appears To Have Violated SB 53 In a Meaningful Way

(14:29) Safeguards They Did Implement

(16:55) Misalignment Risks and Internal Deployment

(18:38) The Official Pitch

(24:28) Inception

(26:12) Turn The Beat Around

(27:35) Codex Does Cool Things

(29:33) Positive Reactions

(38:03) Negative Reactions

(40:43) Codex of Ultimate Vibing

---

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.

Life comes at you increasingly fast. Two months after Claude Opus 4.5 we get a substantial upgrade in Claude Opus 4.6. The same day, we got GPT-5.3-Codex.

That used to be something we’d call remarkably fast. It's probably the new normal, until things get even faster than that. Welcome to recursive self-improvement.

Before those releases, I was using Claude Opus 4.5 and Claude Code for essentially everything interesting, and only using GPT-5.2 and Gemini to fill in the gaps or for narrow specific uses.

GPT-5.3-Codex is restricted to Codex, so this means that for other purposes Anthropic and Claude have only extended the lead. This is the first time in a while that a model got upgraded while it was still my clear daily driver.

Claude also pulled out several other advances to their ecosystem, including fast mode, and expanding Cowork to Windows, while OpenAI gave us an app for Codex.

For fully agentic coding, GPT-5.3-Codex and Claude Opus 4.6 both look like substantial upgrades. Both sides claim they’re better, as you would expect. If you’re serious about your coding and have hard problems, you should try out both, and see what combination works [...]

---

Outline:

(01:55) On Your Marks

(17:35) Official Pitches

(17:56) It Compiles

(21:42) It Exploits

(22:45) It Lets You Catch Them All

(23:16) It Does Not Get Eaten By A Grue

(24:10) It Is Overeager

(25:24) It Builds Things

(27:58) Pro Mode

(28:24) Reactions

(28:36) Positive Reactions

(42:12) Negative Reactions

(50:40) Personality Changes

(56:28) On Writing

(59:11) They Banned Prefilling

(01:00:27) A Note On System Cards In General

(01:01:34) Listen All Yall Its Sabotage

(01:05:00) The Codex of Competition

(01:06:22) The Niche of Gemini

(01:07:55) Choose Your Fighter

(01:12:17) Accelerando

---

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.

Coverage of Claude Opus 4.6 started yesterday with the mundane alignment and model welfare sections of the model card.

Today covers the kinds of safety I think matter most: Sabotage, deception, situational awareness, outside red teaming and most importantly the frontier, catastrophic and existential risks. I think it was correct to release Opus 4.6 as an ASL-3 model, but the process Anthropic uses is breaking down, and it not on track to reliably get the right answer on Opus 5.

Tomorrow I’ll cover benchmarks, reactions and the holistic takeaways and practical implications. I’m still taking it all in, but it seems clear to me that Claude Opus 4.6 is the best model out there and should be your daily driver, with or without Claude Code, on most non-coding tasks, but it is not without its weaknesses, in particular in writing and falling into generating more ‘AI slop’ style prose than Claude Opus 4.5.

For coding tasks, I presume that Opus 4.6 with Claude Code is the play, especially with Agent Teams and fast mode available, and I’m using it myself, but Codex with GPT-5.3-Codex-Max is also a strong model and a viable alternative, and a fully [...]

---

Outline:

(01:32) Sabotage, Deception and Evaluation Integrity

(03:42) Sandbagging On Dangerous Capability Evaluations

(06:01) Situational Awareness

(07:33) Inhibiting Evaluation Awareness (6.5)

(09:06) Self-Preference

(10:24) UK AISI Testing

(11:40) Apollo Research Testing

(14:24) Responsible Scaling Policy Evaluations

(15:45) CBRN (mostly Biology)

(18:43) Autonomy

(26:40) Autonomy Benchmarks

(29:53) Cyber

(31:27) Ship It Anyway

(33:40) You Are Not Ready

---

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.