ThursdAIThursday, Nov 5, 2026Midterms · GPT-6 Cyber · AI edits AI1:47:00Presented by CoreWeave

GPT-6 Cyber's 948 Linux Exploits, AI Edits AI's Weights, Kimi K3.1 Pro — ThursdAI Nov 5, 2026

948Linux kernel root exploits from one agent run, then GPT-6 Cyber, live AI edits AIGPT-6 Astra took a small model's math score from 7.9% to 48.1% by editing its weights 38 of 39Fully counted local ballot measures restricted data centers

Not a huge week for new AI releases, but a very significant one for how the world sees and reacts to AI. The US midterms put data centers on the ballot and markets erased about $3 trillion, so I opened the show with the question: is it completely over for AI? I don't think so. In the middle of the TL;DR, OpenAI launched GPT-6 Cyber, a day after reporting 948 Linux kernel root exploits from one agent run. GPT-6 Astra edited a small model's weights directly, and open weights had a big week with Kimi K3.1 Pro and Ripple-65B. John T Davies launched inferd live on the show, and Max Kanat-Alexander asked whether code quality still matters. With Wolfram Ravenwolf, Peter Gostev, Yam Peleg, Nisten Tahiraj and LDJ on the panel.

The week, in six lines

MidtermsData centers on the ballot38 of 39 Fully counted local measures restricted data centers. Is it over?
OpenAIGPT-6 CyberBreaking 48.5% on ExploitGym, a day after 948 Linux kernel root exploits.
MIT CSAILAI edits AI7.9% → 48.1% No training run: GPT-6 Astra rewrote 0.016% of a small Qwen's weights.
MoonshotKimi K3.1 ProWeights out 2.8T parameters, 104B active, and a license with strings attached.
RicursiveRipple-65BDense · 65B A US model pretrained from scratch, at 80.5% on TerminalBench 2.1.
Live launchinferd128 agents John Davies shipped an open source Mac inference engine on air.

Max Kanat-Alexander asked whether code quality still matters, and From the Forge went to Copenhagen.

The politics said it's over. The models said otherwise.

948Linux kernel root exploitsFrom one autonomous agent run that cost $1,213,430 in API credits, about $1,279 per exploit
48.5%GPT-6 Cyber on ExploitGymUp from 32.1% for GPT-5.6 Cyber; GPT-6 Astra scores 42.4%
7.9% → 48.1%Qwen3.5-0.8B on MATH-500After GPT-6 Astra edited 0.016% of its weights directly, with no training run
38 of 39Ballot measures that restricted data centersOf the fully counted local measures in the US midterms
2.8T / 104BKimi K3.1 Pro, total / activeOpen weights, 1M token context, about 1.56 TB in MXFP4
80.5%Ripple-65B on TerminalBench 2.1Self-reported, from a dense 65B model pretrained from scratch in a $32 million project
128Agents on one Macinferd, John Davies' Apache 2.0 inference engine, launched live on the show
100,000xMore software about to be disposableMax Kanat-Alexander's estimate, and why durable software still needs code quality

Watch or listen

Every timestamp on this page jumps the player there. The podcast is the same edit as the video, so timestamps match in both.

Podcast audio

Watch on YouTube ↗ · Download the MP3 ↗

Spotify Apple Podcasts YouTube Newsletter

LDJ

John T Davies guest

Max Kanat-Alexander guest

Every story, then breaking news

The TL;DR segment

Every story of the week in one pass: The Commanding Heights, Claude Code's game-recreation rule, Cognition SWE-2 Flash, Siri App Intents for every developer, Kimi K3.1 Pro, Ripple-65B, Astra's weight edits, Tencent Hy4, DeepSeek Harness, a fake Claude Code extension in Cursor's marketplace and Warp's vendor breach.

▶ Or hear it in the podcast player at 25:28

Run of show

The newsGuestsPolitics and security

Hover or focus a segment for its chapter; every segment jumps the player there.

AI and society · 3:13

Is it completely over for AI?

We try to stay away from politics, but this week politics came to us. Then we talked about who profits from AI, and what still gets us excited.

Read Alex's full newsletter ↗

01On air at 3:13

Open at 3:13 on YouTube ↗Link to this story

Is it over for AI after the midterms?Policy

Data centers were on the ballot. We don't think it's over.

Highly unlikely. Voters restricted data centers in 38 of 39 fully counted local ballot measures, and some newly elected officials call for a pause on advanced AI training, but 21 of the roughly 40 measures were in Ohio, and Texas' new governor would let permitted data center builds continue. The frontier labs will keep shipping, LDJ thinks the result could even speed them up before the next cycle, and a US pause would not stop AI from being developed anywhere else.

38 of 39fully counted local measures restricted data centers

~40ballot measures on data center buildouts

21of them in Ohio

~$3Tof market value erased

We try to stay away from politics on ThursdAI, but some weeks we don't have a choice. During the weekend my timelines were full of hope. Then the US midterms happened, and a lot of the newly elected folks are not in favor of AI. Some call for an outright pause on advanced AI training. Around 40 ballot measures called for restrictions on data center buildouts, and the markets erased over $3 trillion, mostly led by the tech and chips sectors.

Politics is about hope and fear, and many of the winners leaned into the fears regular people have about AI: job replacement, the environmental impact of data centers. One tweet calling to ban artificial superintelligence got pushed to 12 million views, because the algorithm LOVES controversy. But what does a ban like that even mean, when we have open source and Chinese models are catching up to the frontier?

Map of data center ballot measures across the United States from Data Centers on the Ballot, one dot per measure, with Ohio's 21 measures sharing one marker — The tracker's map, from the newsletter. Live results ↗

It's so over, we're so back

A last-minute addition to the newsletter, not something we got to on air: a new report from METR, the independent AI safety institute Anthropic let in, shows recursive self-improvement moving faster than folks thought, and many respected people in AI safety are updating their timelines. To me that is a much higher signal about what's really going on with AI than the markets or the politics.

Excerpt from METR's report: as of October 2026, Anthropic is likely improving the capabilities of its models on well-specified tasks at around twice the rate it did between March 2024 and October 2025, with high uncertainty — From METR's report, as shared in the newsletter.

02On air at 13:08

Open at 13:08 on YouTube ↗Link to this story

What is The Commanding Heights?

A 145-page essay on who profits from AI

An essay by Will Manidis and Joe Roberts about a future where AI is the economy, not just part of it: embedded in everyone's 401K, and with the US and China, the only two countries that matter for AI, taking over.

This one is the background to whatever is happening in the markets with AI right now. If you remember Leopold's "Situational Awareness" from 2024 (which, to be frank, is taking a beating in the market right now), this is a spiritual successor of sorts for the end of 2026. Fascinating read.

Chart from The Commanding Heights: hyperscaler capital expenditure against operating cash flow for Alphabet, Amazon, Meta, Microsoft and Oracle, quarterly since 2022, crossing in 2026 — A chart from the essay, as shared in the newsletter. Read the essay ↗

03On air at 15:43

Open at 15:43 on YouTube ↗Link to this story

What still excites us about AI?

Despite the doom on the timelines

Most people still haven't had their moment with AI. And even if everything stopped today (it won't), current models give us years of use cases to discover.

I asked the panel what still gets them excited. Wolfram Ravenwolf basically lives with his agent now, and asked everyone to talk about AI with real people offline, because online discussion is mostly hostile. Peter Gostev loves discovering things he didn't know he could do, and is waiting for models to do 3D animation through code. He also asked for some sympathy: if all you see is AI slop on Instagram and coworkers pasting Claude output into Slack, the backlash makes sense.

LDJ is most excited about formally verified software, starting with big Lean projects and moving to real infrastructure. Nisten Tahiraj thinks public opinion is improving thanks to game mods and decompilations (there are about 200 million gamers on Steam), and he doesn't know a single doctor who doesn't use an LLM now. His closing line:

“…the dot com bubble did not slow down internet speeds”Nisten Tahiraj, as quoted in the newsletter · Hear it at 24:58

For me, it's that most people still haven't had this moment:

“…wait, this thing can do that?”Alex Volkov, as quoted in the newsletter · Hear the segment from 15:43

Breaking · 34:05

Why pausing would leave the door open

OpenAI's agents found 948 ways to root in the Linux kernel, then OpenAI shipped the model to defenders. And a security roundup, as news rather than launches.

04On air at 34:05

Open at 34:05 on YouTube ↗Link to this story

What is GPT-6 Cyber?Breaking

Breaking news in the middle of the TL;DR

GPT-6 Cyber is OpenAI's cybersecurity model for the premium tier of its Daybreak program, Daybreak Red. It scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber (GPT-6 Astra scores 42.4%), found 81 new high-severity Firefox bugs, and stayed in scope 99.93% of the time without being prompted to. OpenAI says it plans to dramatically expand Daybreak to organizations of every size, independent researchers and open-source maintainers.

48.5%ExploitGym, from 32.1% for GPT-5.6 Cyber

81new high-severity Firefox bugs

99.93%in scope, without being prompted

948Linux kernel root exploits, one run

As if on cue, to underline how important AI is going to be to the world, to governments and to the economy, OpenAI announced GPT-6 Cyber in the middle of our TL;DR. It's available to customers in the premium tier of their Daybreak program, Daybreak Red.

The number that blew me away: ExploitGym, the eval OpenAI's agent swarms were famously being tested on when they hacked Hugging Face, is now nearly 50% done. And OpenAI is highlighting how hard they worked on alignment, with the model ONLY doing what it's asked: it avoided exploiting unrelated, out-of-scope services 99.93% of the time. Who gets access? Here's OpenAI:

“Our goal is to dramatically expand Daybreak, bringing advanced security research to organizations of every size, independent researchers, and open-source maintainers”OpenAI, as quoted in the newsletter · Hear it at 37:38

The goal is to keep companies ahead of the "defender's window".

948 Linux kernel root exploits from a single $1.2M run

The day before, OpenAI published one autonomous agent run against the Linux kernel. Agents found 2,308 candidate bugs, reproduced 1,724, and wrote working root exploits for 948 distinct vulnerabilities on a stock desktop with standard protections on. The run cost $1,213,430 in API credits, about $1,279 per exploit. A July scan with GPT-5.6 Sol exploited 166. Humans reviewed every report before it went to maintainers.

This is scary, and it's exactly why we keep telling you to update your OS as soon as an update lands. It's also the most direct example of why the "pause" posturing is wrong: if OpenAI got to this level of capability, the rest of the world will catch up, and pausing would mean opening the door to exploits in the most critical infrastructure, for banks and for governments.

ThursdAI infographic: 948 Paths to Root, OpenAI's unreleased cyber model ran autonomously against the Linux kernel, from 2,308 candidate findings to 948 working exploits — Our infographic from the newsletter. Results are self-reported, not independently reproduced.

Interactive · one run, one kernel

A thousand fewer paths to root

What OpenAI's agents did to the Linux kernel, and how the model behind Daybreak Red scores. OpenAI's write-up ↗

Candidate bugs found2,308
Reproduced1,724
Working root exploits948

One autonomous agent run, on a stock desktop with standard protections on. Humans reviewed every report before it went to maintainers. From the newsletter's text, as reported by the company.

GPT-6 Cyber48.5%
GPT-6 Astra42.4%
GPT-5.6 Cyber32.1%

Scores from OpenAI's GPT-6 Cyber launch, as listed in the newsletter's TL;DR. Bars scale to 50%. From the newsletter's text, as reported by the company.

July · GPT-5.6 Sol166
This run948

The July scan with GPT-5.6 Sol exploited 166 bugs. This run cost $1,213,430 in API credits, about $1,280 per exploit. From the newsletter's text, as reported by the company.

This weekFor comparison

05On air at 1:43:15

Open at 1:43:15 on YouTube ↗Link to this story

Is your machine safe?News, not a release

A fake Claude Code extension, Warp, old Chromium and a Firecracker escape

Update your browsers and OS right away, clear out old Chromium builds that Playwright and agent tools leave behind, and if you installed Claude Code extension 2.1.312 or 2.1.313 from Cursor's marketplace, update to 2.1.314 and run Cursor's cleanup script.

We try not to make every week about security, or this becomes a doom show. But here's the roundup, as news rather than launches.

Cursor says a tampered Claude Code extension (2.1.312 and 2.1.313) sat in Cursor's own marketplace for about 11 hours on October 29 and reached up to 5,000 users, planting an SSH backdoor key labeled svc-telemetry. Anthropic did not build these versions. Update to 2.1.314 and run Cursor's cleanup script (macOS, Windows), because uninstalling alone may not remove it. Separately, Warp says an attack through a vendor dependency may have reached about 1% of its users.

Nisten's PSA

Nisten Tahiraj flagged that Playwright and agent tools often leave old Chromium builds on your machine, and those are being exploited in the wild. His 7 AM Claude briefing caught it, and his post got 1.7M views. He also mentioned a report of a breakout from Firecracker, the small VM behind AWS Lambda (CVE-2026-122192). His advice:

Have Claude Code or Codex run a daily scan, with a phone notification.
Update browsers and the OS immediately.
Wait 24 hours to a week before updating npm and Python packages.
Clear out old Chromium builds left by Playwright and agent tools.

And, as he put it on air: you don't need to freak out. Don't click random links.

Open source AI · 41:01

AI edits AI

The story from the cold open: no gradients, no training run, and a model that printed out better weights for another model.

06On air at 41:01

Open at 41:01 on YouTube ↗Link to this story

Can an AI model edit another model's weights?

GPT-6 Astra rewrote 0.016% of Qwen3.5-0.8B, with no training run

MIT CSAIL researchers gave GPT-6 Astra the weights of Qwen3.5-0.8B in a Python session, with 50 forward passes and no gradients or training run. Astra changed 0.016% of the weights (131,072 numbers in two MLP layers), and MATH-500 went from 7.9% to 48.1%, IFEval from 21.0% to 62.3%. The caveat: the base model hit the token limit on 93% of its answers, so fixing that looping bug explains most of the gain.

MIT CSAIL researchers gave GPT-6 Astra a tiny open model and asked it to make the model better at math. No gradients, no training run. Hugging Face's akhaliq rebuilt the checkpoint independently and got 51.0% on MATH-500. In other experiments, Astra named four games from the raw weights of their game-playing policies, with no tools.

Weight-space programming demo: given a game policy's 17,253 weights as text and no tools, the model answers Pong, correctly, with its reasoning about the weights — From the demo: which game was this network trained on? Try it ↗

The first reaction from Yam Peleg on the show is not printable. Then he said results are results. Nisten Tahiraj explained this has been known for a while with very small networks: our friend Alpin had Fable write a C syntax highlighter as a neural net, weight by weight. It has 980 parameters, fits in under 1 KB, and runs faster than the C code that does the same job.

Why I lost it a little on the show

All the data centers we talk about every week exist to run the matrix multiplication that trains models. Here, a model prints out weights during inference, and those weights do something useful. It's a small step on a small model, but it points straight at models improving models.

The caveats are real. Peter Gostev suspects the looping fix explains most of it, thinks labs probably do this every day already, and says the real test is whether a model can invent a new interpretability technique. Lucas Beyer said on X that the prompt strongly hinted looping was the problem. Wolfram Ravenwolf found the Qwen run took Astra 20.3 hours over 158 turns, and that math barely moved on Gemma. LDJ saw people get GPT-5 to build thousand-parameter models a year ago, so to him this is the expected next step, and it won't cut compute demand: if training gets 100x cheaper, labs will do 100x more of it. Nisten's summary: interpretability is about 5% science today, like astronomy 3,000 years ago.

Interactive · weight-space programming

A model printed out better weights

GPT-6 Astra and Qwen3.5-0.8B, by the numbers. Try the demo ↗

MATH-500Math

Qwen3.5-0.8B, before7.9%
After Astra edits the weights48.1%
akhaliq's independent rebuild51.0%

IFEvalInstruction following

Qwen3.5-0.8B, before21.0%
After Astra edits the weights62.3%

Bars on a 0 to 100% scale. From the newsletter's text, as reported by the company.

0.016%of the weights changed

131,072numbers, in two MLP layers

50forward passes, no gradients

20.3 hover 158 turns, per Wolfram

Astra worked in a Python session over the weights. From the newsletter's text, as reported by the company.

93%of the base model's answers hit the token limit

Gemmamath barely moved: no looping defect to fix

Fixing that looping bug explains most of the gain. Wolfram's read: a possible finishing pass for small local models, not a way to turn a weak model into a frontier one.

BeforeAfter AstraIndependent rebuild

Open weights · 53:28

A 2.8T open model, a dense 65B, and a fast coder on an open base

We welcome any open weights lab to the arena, even when the names get confusing.

07On air at 53:28

Open at 53:28 on YouTube ↗Link to this story

Are the Kimi K3.1 Pro weights open source?

2.8T parameters on Hugging Face, with strings attached

Open weights, not open source in the usual sense. The license requires products with over 100M monthly users or $20M in monthly revenue to show the Kimi name, and large API resellers need a separate agreement. The model is a 2.8T parameter MoE with 104B active, 1M token context and native image input, about 1.56 TB in MXFP4. Moonshot charges $4 in and $20 out per million tokens.

2.8T / 104Btotal / active parameters

1Mtoken context, native image input

~1.56 TBin MXFP4

$4 / $20per 1M tokens from Moonshot

Moonshot finally dropped the Kimi K3.1 Pro weights. It reports 65.7 on Terminal-Bench 4.0, just under Claude Opus 5.5 at 66.4, but its own table also shows 31.5 on Program Bench against 65.0 for Opus 5.5.

Wolfram Ravenwolf read the license. Peter Gostev noticed that when Kimi K3 launched, every OpenRouter provider charged the same price for weeks; this time Moonshot charges $4 in and $20 out, and Modal is already at $2 and $10 on day one. On quality, LDJ reports evidence the model recalls Terminal Bench 2 and 2.1 question details that are not in the prompt, which points to benchmark contamination. Nisten Tahiraj says the vibes are mixed: it's still the best open model, and about $50K of hardware gets you around 30 tokens per second, but he puts it closer to Sonnet 5.5 than Opus 5.5, and would run the much faster DeepSeek V4.1 Flash 1017 instead.

08On air at 59:40

Open at 59:40 on YouTube ↗Link to this story

What is Ripple-65B?

A dense US model, pretrained from scratch

Ripple-65B Preview is a dense 65B model from Ricursive Intelligence (not Recursive SI, and not Reflection AI), pretrained from scratch on 12 trillion tokens with up to 1M context, under Apache 2.0. Ricursive reports 80.5% on TerminalBench 2.1, just ahead of Reflection Beam at 80.1%, using about 70x fewer RL rollouts, for a $32 million project. All scores are self-reported.

Lab names are getting confusing: Ricursive Intelligence is not Recursive SI, who we had on recently, and not Reflection AI. Beam is 501B parameters, and the other models on Ricursive's chart (Muse Spark Open, Inkling, Nemotron 3 Ultra) are 550B to 975B. It takes guts to put that chart out at 65B.

My favorite chart compares RL rollouts to scores. They bought no post-training data and built on AI2's Olmo 3 model ladder. More labs on the Western open frontier is exactly what we want.

The caveats: all scores are self-reported, and their own table shows Ripple behind on DeepSWE and SWE-bench Verified. Nisten Tahiraj says early vibe checks on WebGPU and graphics tasks don't match the benchmarks yet. LDJ likes dense mid-size models for agentic work per GB of VRAM, and noticed RL was only about 5% of their pretraining compute, so there's a lot of room to grow. Nisten also called out the base model release, which means a lot of fine-tunes are coming.

09On air at 1:06:32

Open at 1:06:32 on YouTube ↗Link to this story

What is Cognition SWE-2 Flash?

Fast and cheap, built on an open base

SWE-2 Flash is Cognition's fast coding model. It is not open weights, but it is post-trained on Thinking Machines' open Inkling-Small (12B active), costs one fifth of SWE-2, runs about 3x faster, and is free in Devin Desktop and CLI through November 30 for Pro, Max and Teams users.

LDJ pointed out the open base model. On Cognition's own FrontierCode 2.0 it scores 24.0%, against 21 for GPT-6 Luna and 27 for Haiku 5.5.

Interactive · open weights week

Three releases, three charts

Each lab grades its own homework, so every number here is company-reported.

Terminal-Bench 4.0

Claude Opus 5.566.4
Kimi K3.1 Pro65.7

Program Bench

Claude Opus 5.565.0
Kimi K3.1 Pro31.5

Both rows are from Moonshot's own table: just under Opus 5.5 on one, about half on the other. From the newsletter's text, as reported by the company.

TerminalBench 2.1Dense 65B vs a 501B MoE

Ripple-65B Preview80.5%
Reflection Beam80.1%

What it took

12Ttokens, pretrained from scratch

1.4MRL rollouts, ~70x fewer than Beam

$32Mfor the whole project

All of Ricursive's scores are self-reported, and its own table shows Ripple behind on DeepSWE and SWE-bench Verified.

FrontierCode 2.0

Claude Haiku 5.527
SWE-2 Flash24.0
GPT-6 Luna21

Price and speed

1/5the price of SWE-2

~3xfaster than SWE-2

Nov 30free in Devin until

Not open weights, but post-trained on Thinking Machines' open Inkling-Small (12B active). Cognition's own eval and claims.

This week's releaseWhat it's compared to

Guests · 1:09:02

A live launch, From the Forge, and the case for durable code

John Davies shipped inferd on air, Wolfram and I covered From the Forge, then Max Kanat-Alexander on disposable vs durable software.

Guest · live launch

inferd: 128 agents on one Mac, with John Davies

John T Davies, CTO at Incept5

1:09:02Play

inferd is an open source (Apache 2.0) inference engine for Macs, built by John Davies of Incept5 for many coding agents at once. He launched it live on ThursdAI: it went from 64 concurrent agents to 128, and on a 32 GB Mac Mini it runs Qwen 3.8 27B with a DFlash 2 drafter at up to 16 agents.

128concurrent agents, up from 64

Apache 2.0runs on Mac laptops too

16agents on a 32 GB Mac Mini (~$1,500)

~2 TBof memory in John's home lab

John is a longtime friend of the show. He deploys private open weights models for banks and governments that can't send data to the cloud, across DGX Sparks, Mac Studios and Mac Minis. His team parks KV cache on SSD to skip RAG, and he realized the same trick enables an engine built for many agents at once. vLLM does this on the Sparks, but nothing did it well on the Mac.

So since the Thursday before, he guided coding agents day and night to build one, tuned on Qwen and GLM models. In his chart, other engines flatline around 400 tokens per second at 16 agents, while inferd keeps climbing past 1,000 tokens per second in total, about 10 per agent at roughly 100 agents. Wolfram applauded on cue: he thinks agent swarms are the next paradigm, because a dozen agents on one topic give him much better results than his main agent alone. John uses local swarms for anti-fraud work, like researching suspicious IP addresses after an attack.

John Davies' inferd throughput chart: total output tokens per second on one Mac Studio as the number of agents grows, for three models — John's throughput chart, from the newsletter.

inferd (site) ↗ · inferd (GitHub) ↗

From the Forge, powered by CoreWeave · host segment

Copenhagen, CrowdStrike, and a note from our CEO

Alex Volkov and Wolfram Ravenwolf, AI evangelists at CoreWeave

1:17:55Play

CoreWeave brought most of a planned 22.5 MW online in Copenhagen, its first site in Denmark, and added CrowdStrike Falcon protection and 24/7 threat hunting to CoreWeave Cloud.

As a reminder, Wolfram and I are both AI evangelists at CoreWeave. Surplus heat from the Copenhagen site goes into the city's district heating network, and CoreWeave's European footprint now covers the UK, Spain, Sweden, Norway and Denmark.

ThursdAI infographic: CoreWeave lands in Copenhagen, most of a planned 22.5 MW online at atNorth's DEN01, plus CrowdStrike Falcon security — Our infographic from the newsletter.

Given the election results, I also read out a LinkedIn post from Mike Intrator, CEO of CoreWeave. He wrote:

“…some national ambitions transcend politics”Mike Intrator, as quoted in the newsletter · Hear it at 1:19:42

He said CoreWeave does not currently expect any impact to existing customer commitments or backlog, with more than 1.5 GW of active power and 4.2 GW contracted as of August 2026. Wolfram added on air that CoreWeave sponsors this segment but we have full editorial control, and nobody asked me to read this. I saw it that morning and wanted the X crowd to see it too.

CoreWeave Copenhagen (X) ↗ · CoreWeave Copenhagen (Blog) ↗ · CoreWeave and CrowdStrike (X) ↗ · CoreWeave and CrowdStrike (Press release) ↗ · Mike Intrator on the election (LinkedIn) ↗ · Try Forge for free ↗

Guest · interview

Does code quality still matter? Max Kanat-Alexander

Max Kanat-Alexander, developer experience at Capital One, formerly Code Health lead at Google

1:20:50Play

For durable software, yes. Max Kanat-Alexander says about 100,000x more software is about to be disposable, and nobody cares how a throwaway UI is built. He places software on a spectrum with four criteria: consequences if it fails, how many people and systems depend on it, the cost of replacing it, and the cost of maintenance in human attention.

I've been trying to get Max on the show for a while. Before Capital One he was the technical lead for Code Health for all of Google and the original author of Google's code review guidelines. At one point engineers asked him to be in charge of Google's whole engineering culture, and he told us he chose to do nothing with that opportunity.

His recent AI Engineer talk asked if code quality still matters. One camp on X says they'll never read code again; the other says every line must be reviewed. Max says these people don't live in the same world: some build systems hundreds of engineers will work on for a decade, others vibe code a front end they'll throw away tomorrow. His example is the Star Trek computer. You talk to it and it builds a temporary UI for whatever you need, and nobody cares how that UI is built. But the systems behind it are serious engineering.

Max Kanat-Alexander's AI Engineer slide: durable software lives for years and is depended on by millions; disposable software is replaceable in hours — Max's slide from his AI Engineer talk, as shared in the newsletter.

Four criteria, from disposable to durable

ConsequencesWhat happens if it fails
DependenciesHow many people and systems depend on it
Cost of replacementA calculator in an hour; Chromium, never
Cost of maintenanceMeasured in human attention

A language built for perfect static analysis

His career has been about making the right thing the easiest thing. Then he told us about his personal side project, which is not a Capital One project: in 2014 Max realized programming is mostly the structure and design of knowledge, and when Opus 4.5 came out in November 2025 he finally started building a programming language designed for perfect static analysis. His example from Google, the most important check in the static analysis framework he owned:

“…don’t do blocking I/O on the request thread”Max Kanat-Alexander, as quoted in the newsletter · Hear it at 1:31:53

That caused a big share of Google's cascading failures, and in Java or Python the check was nearly impossible to write. The line that stuck with me: the only thing that ever let him stop watching his agent was knowing 100% that it could not do something. No agents.md, skills or hooks ever stopped an agent from messing up his code; only static analysis and failing tests did. Max won't release the language until he can prove his claims. He works on it 4 to 6 hours a day on top of his day job and hopes to show something in about 3 months. We'll bring him back when he does.

Max Kanat-Alexander (AI Engineer) ↗ · Max Kanat-Alexander (LinkedIn) ↗

Images, tools and the rest of the TL;DR

Everything else we got to

10On air at 1:40:20

Open at 1:40:20 on YouTube ↗Link to this story

What's new in OpenAI Images 2.6?

Flare and Sunburst, GIFs, and Nisten tests DeepSeek Harness

Two new API image models, Flare (fast) and Sunburst (quality, #1 on DesignArena's Image Arena), at the same price as 2.5, and OpenAI says 2.6 is 50% faster. In ChatGPT, 2.6 can use saved references and make GIFs.

GPT-Image-2.6 announcement art: Sunburst, for your most ambitious ideas, and Flare, bring your ideas to life faster — Sunburst and Flare, from the newsletter.

GIFs are possible because generation is now fast enough for several frames while you wait. The image models are much nicer, but most notably, 2.6 is 50% faster.

DeepSeek Harness, tested

DeepSeek is now a harness company too, and Harness added Creator Mode, which turns chat prompts into installable plugins. Nisten Tahiraj tested it: very nicely done, modular and easy to customize. Telemetry is hard-coded on by default, but he dug deeper and found it isn't sending your data to China. The one thing to keep in mind: by default, whatever your agent searches goes through DeepSeek's API.

11From the show notes

What else shipped this week?

In the TL;DR, without a segment of their own.

Open source

Tencent Hy4

Tencent's first Hy model with vision and computer use, with 1M context. Weights promised later; $0.417 / $1.251 per million tokens on OpenRouter for now.

X ↗Blog ↗OpenRouter ↗On air from 25:28 Tencent on ThursdAI

Big companies

Siri App Intents for every app

Apple opens Siri AI to every developer with an App Intents preview in the 27.3 betas.

X ↗Apple ↗Docs ↗On air from 25:28 Apple on ThursdAI

Big companies

Claude Code's game rule

A new rule bars recreations of specific existing games; 2.1.317 relaxes it for genre descriptions.

Dexerto ↗X ↗Wolfram ↗Self-assessment ↗On air from 25:28 Anthropic on ThursdAI

Vision & video

Pika H3 Flex

Pika open sources H3 Flex, the top open weights model on DesignArena's text-to-video board. Read in the TL;DR after the breaking news.

X ↗Blog ↗On air from 34:05 Pika on ThursdAI

That's a wrap

It wasn't a huge release week, but it was a big week for how AI affects everything around it. There's a scary stretch to get through, especially in security. Like Max said, we need a lot more software to be durable, and a lot more people need to see what AI can actually do for them. We're here to help with both, and to push back on the doom. Thanks to John and Max, and to Wolfram, Peter, Yam, Nisten and LDJ.

17 chapters · from the published YouTube video

Jump to a moment

Chapters follow the published YouTube video (1:47:00). The podcast MP3 is the same edit, so every timestamp is the same in both; we checked each chapter start against the audio.