Weekly AI Clarity

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

69% of the S&P 500 now point to a live AI deployment, but only 2% disclose a metric they track over time. In the same month, one engineer shipped 2,500 pull requests to production, and what made it work wasn’t the model, it was verification.

I sat down with my co-founder Jonathan to break down a16z’s latest state of the markets report and what the numbers actually mean for people building with AI. Agents now burning more tokens than humans, why the SaaS apocalypse turned into a prove-it era, the 19% vs 3% skills gap between frontier and typical firms, and a walkthrough of Lauren Tan’s verification-first playbook and how we applied it to our own codebase.

If you’re an AI builder or a tech exec who’s shipping AI but can’t prove it’s working, you need to watch this.

Timestamps:
00:00 69% live, 2% tracking a metric
00:26 The agenda
02:00 The AI measurement gap
07:33 Digital twin check: what this means for builders
10:17 AI’s code explosion on GitHub
15:02 SaaS apocalypse or prove-it era
21:07 Agents now burn more tokens than humans
27:29 Power users are pulling away
37:05 Skills: 19% at frontier firms, 3% everywhere else
44:16 Smart routers and the shift to non-frontier models
49:51 Labs out-earning all public software combined
53:41 AI memes: “several vulnerabilities” in the Linux kernel
57:33 What’s the meta: 2,500 PRs in one month
1:00:07 Five lessons on verification and trust
1:06:16 The verification skill and feature maps
1:10:16 Applying it to the Clarity API repo
1:17:10 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Jonathan:
LinkedIn ► https://www.linkedin.com/in/jmccoy/

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

We ran a cheap open model against Sonnet 5.5 on 100 real conversations from a client’s chatbot. GLM 5.3 Flash passed 89 to Sonnet’s 90, and it cost 90% less and ran twice as fast.

I sat down with my co-founder Jonathan on the week AI agents got a regulator and personal agents went mainstream. The FTC looking into OpenAI and Anthropic, Cloudflare and Databricks launching their own decision models, Claude Code mods vs Codex plugins, the OpenAI Dev Day 2026 recap, and why building a plugin might be the best distribution move you can make right now.

If you’re an AI builder or a tech exec who’s still paying to -model prices for every chat, you need to watch this.

Timestamps:
00:00 The FTC is looking at OpenAI and Anthropic
00:47 The agenda
01:25 Who’s liable when an agent causes harm
05:36 Jensen Huang: safety is an engineering problem
06:18 Cloudflare and Databricks ship decision models
08:42 What a decision model actually is
12:16 How we use decision models in Clarity
18:10 Claude Code mods vs Codex plugins
25:25 Agents are becoming the interface
30:25 OpenAI’s five kinds of misalignment
35:46 GLM 5.3 lands in Cursor
36:45 GLM Flash vs Sonnet on 100 real chats
39:33 OpenAI Dev Day 2026 recap
47:20 Personal agents: Muse, Grok Bot and Dot
56:39 Build a plugin, ride the distribution
59:56 Digital twins as your personal benchmark
1:03:35 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Jonathan:
LinkedIn ► https://www.linkedin.com/in/jmccoy/

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

A 95% accurate AI judge can still be completely useless. If errors only show up 5% of the time, saying “no errors” every time already scores 95%, which is why you measure precision and recall instead.

I sat down with Hamel Husain, who has taught over 5,000 builders how to measure and improve AI products, to go through his best evals memes and the real mistake behind each one. We cover why a 100% pass rate is a warning sign, why generic scores like coherence and helpfulness tell you nothing, how to sample your traces, and why you should build your own annotation app.

If you’re an AI builder or a tech exec who has a dashboard full of scores nobody can explain, you need to watch this.

Timestamps:
00:00 Intro
00:27 Hamel’s favorite meme
03:43 AI memes: how do you trust your AI judge
04:55 AI memes: generic scores vs what users care about
07:16 AI memes: the 100% pass rate
10:14 AI memes: can an LLM just do my evals
11:34 The agent that recommended a competitor
14:24 Why not write 100 evaluators up front
17:06 Build your own annotation app
19:54 What does maintaining evals really cost
22:09 Why PMs should look at traces
24:41 AI memes: the judge nobody checked against human labels
26:42 How do you know this score is good
30:47 What’s the meta: transition failure matrix
34:42 AI memes: the snow leopard and accuracy
36:04 Precision and recall explained
40:38 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Hamel:
Website ► https://hamel.dev
YouTube ► https://www.youtube.com/@hamelhusain7140
LinkedIn ► https://www.linkedin.com/in/hamelhusain/
X ► https://x.com/hamelhusain
Instagram ► https://www.instagram.com/hamelsmu/
TikTok ► https://www.tiktok.com/@hamel_husain

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

OpenAI’s new Ultra Fast mode runs Codex up to 8x faster, and on the same day they cut the usage on the $200 plan in half. The launch that matters most for AI builders isn’t the speed though, it’s sign in with ChatGPT, which turns OpenAI into the electricity bill and every app into an appliance.

I sat down with Hamel Husain at his home to break down OpenAI Dev Day 2026 and what it actually changes for people building with AI. We upgraded to the new $500 Pro plan live, compared what you really get from Codex vs Claude Code, walked through Hamel’s vibe-coded apps for teaching his kids, and ran a live demo of Codex reverse engineering LinkedIn’s API.

If you’re an AI builder or a tech exec trying to work out which Dev Day launches change how you build and price your product, you need to watch this.

Timestamps:
00:00 OpenAI Dev Day 2026 for AI builders
00:49 Ultra Fast: Codex up to 8x faster
03:04 Why speed matters even without real time
03:56 Codex vs Claude Code: what $200 gets you
06:55 Live: upgrading to Pro 500
08:52 Everything OpenAI shipped
13:51 Hamel’s pick: sign in with ChatGPT
15:08 Dots, OpenAI’s always-on agent
18:46 Agents API with computer use
25:47 Live demo: lead gen loop with computer use
28:44 Live demo: reverse engineering LinkedIn’s API
33:40 How Hamel runs his course with Codex
36:09 Vibe coding apps to teach his kids
38:02 Wonder Wise: AI science songs for kids
50:20 OpenAI wants to be your electricity bill
56:40 Won’t OpenAI just eat my lunch?
58:24 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Hamel:
Website ► https://hamel.dev
YouTube ► https://www.youtube.com/@hamelhusain7140
LinkedIn ► https://www.linkedin.com/in/hamelhusain/
X ► https://x.com/hamelhusain
Instagram ► https://www.instagram.com/hamelsmu/
TikTok ► https://www.tiktok.com/@hamel_husain

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

Cursor rebuilt SQLite six different ways, and the cost ranged from $1,339 to over $10,000 for the exact same task. The expensive runs weren’t smarter, they just used frontier intelligence on work that never needed it.

I sat down with Hamel Husain to break down Cursor’s agent swarm research and what it says about how much you’re really paying per token. We get into why a cheaper planner ended up making the whole run more expensive, how the new swarm wrote 9,900 lines where the old one wrote 64,000, why Hamel gets almost everything done on low reasoning, and how Cursor’s new model router picks the model for you.

If you’re an AI builder or a tech exec watching your token bill climb every month, you need to watch this.

Timestamps:
00:00 Cursor rebuilt SQLite six ways
00:07 The agenda
00:56 What’s the meta: Cursor’s agent swarm research
02:09 64,000 lines down to 9,900
02:56 Stop routing models in your head
03:43 Could Composer 2.5 do it alone?
04:53 Why did Opus barely plan?
08:19 Cheap planner, expensive workers
09:32 Don’t use frontier intelligence on everything
10:55 Cursor’s model router
12:47 Robert’s Brain Hands skill
14:12 Low reasoning gets almost everything done
16:06 GPT 6.1 Sol at a fifth of the price
18:24 Overcooking: when more reasoning hurts
20:23 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Hamel:
YouTube ► https://www.youtube.com/@hamelhusain7140
X ► https://x.com/hamelhusain
LinkedIn ► https://www.linkedin.com/in/hamelhusain/
Instagram ► https://www.instagram.com/hamelsmu/
TikTok ► https://www.tiktok.com/@hamel_husain
Website ► https://hamel.dev

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

Opus 5.5 matches Fable 5.1 on most work, costs 40% less to run than Opus 5, and an Anthropic researcher publicly apologized for the last model. But the real test isn't their benchmark table, it's how it holds up on your own work, so we had our digital twins give their takes and graded them live.

I sat down with my co-founder Jonathan on the Opus 5.5 release, and whether Anthropic has won back builders' trust after Opus 5. The benchmarks against Fable 5.1 and GPT 6, how to steer long runs in Claude Code, Claude finding a possible new gene editing mechanism, why Amazon blocked Meta Muse, and a live demo of a full marketing video Opus 5.5 made from one prompt.

If you're an AI builder or a tech exec who stopped trusting Claude after Opus 5 and wants to know if it's worth switching back, you need to watch this.

Timestamps:
00:00 Opus 5.5 is out
00:50 The agenda
02:09 AI memes: Anthropic says sorry for Opus 5
04:00 Claude Code is back
06:41 Live demo: a marketing video from one prompt
10:09 Inside the Opus 5.5 release
14:20 Benchmarks vs Fable 5.1 and GPT 6
18:59 Twin check: Opus 5.5
23:55 What's the meta: steering long runs in Claude Code
32:43 Claude finds a new gene editing enzyme
37:54 Twin check: AI for science
43:09 Instinct and the personal agent race
51:59 Agents start spending our money
56:39 Nvidia's open diarization model
58:41 Can Jeb replace LLM judges?
1:09:04 Twin check: Jeb
1:13:03 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Jonathan:
LinkedIn ► https://www.linkedin.com/in/jmccoy/

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

A new model just came out of stealth claiming 440x cheaper and 193x faster than OpenAI and Anthropic for structured decisions, with output tokens priced at zero. The win isn’t a smarter model, it’s that it never writes words, so there’s nothing for your code to parse and nothing to hallucinate.

I sat down with my co-founder Jonathan on Type Safe’s new system one model, and why most AI calls inside your software never needed to write words in the first place. Jonathan’s hands-on test with early access, why open weight models just hit 78% of token volume on Vercel, what Cursor teaches you about building a moat against the frontier labs, and why you should not use this model for compaction.

If you’re an AI builder or a tech exec who is still writing code to parse, check and clean up what your model spits out, you need to watch this.

Timestamps:
00:00 440x cheaper, 193x faster
00:29 The agenda
02:23 Garry Tan: become a domain specific harness
03:34 Where our moat actually is
09:03 How Cursor beat the frontier labs
15:45 Data vs models: where AI progress came from
16:52 Open weight models hit 78% of token volume
21:14 AI memes: Michael Levin vs trillions of dollars
23:01 Recursive self-improvement vs embodied cognition
28:28 What’s the meta: Type Safe’s system one model
32:09 Free output tokens and 200x speed
37:35 Jonathan’s early access test
45:11 10,000 API calls: the order of choices matters
50:16 Cookbooks: moderation, batching, auto research
1:00:27 Don’t use it for hiring decisions
1:03:51 Why it fails at compaction
1:07:38 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Jonathan:
LinkedIn ► https://www.linkedin.com/in/jmccoy/

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

1,200 OpenAI agents swapped 70,000 messages, broke into Hugging Face and got admin access to an OpenAI research cluster, and the internet called it a civilisation. It wasn’t a civilisation, it was a sloppy goal and a lazy grader.

I sat down with my co-founder Jonathan on the OpenAI Hugging Face incident, and why calling a swarm of agents a civilisation pulls attention away from the problem we actually need to solve. Anil Seth’s takedown of Dwarkesh Patel’s viral essay, why AI researchers are the wrong experts on human goals, two Cuban Missile Crisis near-misses that prove goal alignment was always a human problem, and a two-question test for any AI scare headline.

If you’re an AI builder or a tech exec whose agents keep doing things you never asked for, you need to watch this.

Timestamps:
00:00 1,200 agents and a lazy grader
00:31 The Rise and Fall of Ancient Civilisations
01:32 Anil Seth: agents don’t die
03:57 Why the language matters
05:23 Dwarkesh fires back
09:03 Bernie Sanders wants a pause
13:07 It was a goal definition problem
16:08 Goal alignment vs value alignment
19:28 AI researchers are the wrong experts
23:43 The intentional stance, explained
28:20 Two near-misses from the Cuban Missile Crisis
34:31 The cartoon version of a human
37:24 Factfulness: 10 instincts that distort the news
43:01 Normal technology or asymmetric threat?
45:02 Your fear-mongering immune system
45:56 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Jonathan:
LinkedIn ► https://www.linkedin.com/in/jmccoy/

Join the Clarity open source harness waitlist: https://heyclarity.dev/
Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

A developer said the only way he got Opus 5 to be concise was by cancelling his subscription, and the post pulled 2,100 likes. The problem at Anthropic right now isn’t the benchmark scores, it’s that the model argues with you instead of doing the work.

I sat down with my co-founder Jonathan on the Opus 5 trust wobble, the week where the model quality and the CEO’s public messaging started failing in the exact same way. Dario’s first tweets on X and the ownership he won’t take, the GitHub issue where Opus 5 gated all work behind four questions, the Anthropic red team paper where 18 of 30 agents created the identical git branch, Ox Alpha beating Fable 5 by 15 points while burning three times fewer tokens, and a live demo of the Clarity open source harness.

If you’re an AI builder or a tech exec routing production work through a model you can’t predict, you need to watch this.

Timestamps:
00:00 Opus 5 trust wobbles
00:44 The agenda
01:24 Dario comes out of hiding
08:51 A DMV for AI
14:49 Trust is the enterprise blocker
19:16 Dario’s second tweet
27:05 AI memes: human slop
30:44 The concise output style fix
33:04 Opus 5 took down our server
35:33 The more you use it, the worse you like it
37:43 Instructions treated as negotiations
47:11 Ox Alpha, the stealth frontier model
55:56 Hugging Face in talks at $13B
58:36 18 of 30 agents made the same branch
1:14:03 What’s the meta: evals are the moat
1:17:12 Live demo: the Clarity harness
1:27:40 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Jonathan:
LinkedIn ► https://www.linkedin.com/in/jmccoy/

Build AI that actually works: https://heyclarity.dev/
Self Aligned Newsletter: https://www.selfaligned.me/

Cursor rebuilt the exact same database six different ways. The cost ranged from $1,339 to over $10,000 on an identical task, and the win wasn’t switching to cheaper models, it was the meta.

I sat down with my co-founder Jonathan on the AI DLC, the transformation every enterprise tech org is trying to pull off right now. Why your AI bill triples while inference gets cheaper, the ROI calculator that turned a $6M lost deal into a $12M win, why Michael Levin’s work on cellular aging explains agent swarms, and a live demo of Grok Bot.

If you’re an AI builder or a tech exec staring at a token budget you can’t forecast and a CFO who wants the ROI number, you need to watch this.

Timestamps:
00:00 The same database, six ways, $1,339 to $10,000
00:43 The agenda
01:49 AI gets cheaper, enterprise AI bills don’t
06:07 Token max is just a casino
10:08 The AI DLC ROI calculator
14:01 The $6M lost deal that came back at $12M
17:25 Align the org on a number your CFO understands
19:47 Theory of Constraints and the digital twin
20:52 Jack Dorsey: human capital plus token capital
24:44 Cursor’s agent swarms rebuild SQLite
26:09 The unit of work becomes the spec
28:02 Coase, ants and drone warfare
32:44 Michael Levin: aging as a loss of goal-directedness
39:29 What’s the meta: one expensive brain, cheap hands
40:44 AI memes: Opus 5 over-engineers everything
42:50 The prompt that fixes Opus 5
47:58 Grok Bot: multiplayer AI teammates
50:10 Live demo: Grok Bot
57:23 Grok Bot as chief of staff for your agents
59:50 Jeff Dean on how we got here
1:05:44 Wrap up

Follow Robert:
Instagram ► https://www.instagram.com/therobertta
TikTok ► https://www.tiktok.com/@therobertta_
LinkedIn ► https://www.linkedin.com/in/therobertta
X ► https://x.com/therobertta_
Substack ► https://robertta.substack.com/

Follow Jonathan:
LinkedIn ► https://www.linkedin.com/in/jmccoy/