It's Only Logical

Six days after Matthew predicted a major hack, Binance lost more than $6 billion in crypto. Dara plays the tape. Matthew would like some credit, although he denies any involvement.

Then it’s back to the less reliable business of predicting AI. Dara and Matthew compare the models they’re using, debate how much more intelligence they actually need, and follow the money behind the latest releases. Can Claude be worth every penny to its users while Anthropic is worth less than investors think? They get into the IPO filing, vast compute commitments, and what AI answers could do to Google’s search ads.

Along the way, Dara gives Codex control of his phone, Matthew builds a small app for his family, and both hosts try to work out which improvements make a difference outside a benchmark. They close with bets on the next Claude and OpenAI releases and whether Dot will reach Europe by the next episode.

Follow It’s Only Logical in your podcast app, or subscribe on YouTube, for the next episode.

Chapters

00:00 It’s Only Logical

00:15 Matthew called the hack. Play the tape.

06:35 The fizz check and when to walk away from a task

12:52 Giving Codex control of a phone

17:50 OpenAI’s 28 days of updates and banked resets

24:09 Could Claude change without the label changing?

27:39 GPT-6.1 Terra and a benchmark worth sticking with

34:37 Gemini 4 Carbon, Argon and Google’s model names

38:32 What AI answers do to Google’s search ads

43:16 Anthropic’s IPO and the $895 billion compute bet

49:00 How much more intelligence do we actually need?

54:04 What Anthropic says about AI improving AI

58:25 What does the average person want AI to do?

1:02:17 HackerTube and software built for one family

1:06:59 Predictions for Claude, OpenAI and Dot in Europe

1:10:42 A dedication to Patrick Stewart

Links and on-air follow-ups

- Matthew’s prediction and the FBI account:

https://youtu.be/AKXzS6CfLso?t=5502

https://www.fbi.gov/investigate/cyber/alerts/2026/north-korea-responsible-for-6-billion-binance-attack

- Claude rumour: this likely source discusses possible stronger-model routing under Fable 5.1; Dara recalled possibly Opus 5.5. The underlying model is unconfirmed.

https://www.youtube.com/watch?v=Tyoyvspkcfc&t=18s

https://lesscarbs.net/news/anthropic-fable-rumors-arena-november-briefing/

- Artificial Analysis: live scoreboard/methodology, not recording-date scores (v5.0 then; v5.0.1 now).

https://artificialanalysis.ai/leaderboards/models

https://artificialanalysis.ai/methodology/intelligence-benchmarking

- Likely tracker mentioned: BridgeBench Nerf Bench. Opus 5.5’s 105.8% (16 October) is within 90–110% normal variance. Its score combines correctness/efficiency, not proof of greater intelligence.

https://www.bridgebench.ai/nerf-bench

https://www.bridgebench.ai/blog/the-methodology-of-nerfbench

- Google: Q3 2026 Search & other ad revenue grew 12% year on year; all Google advertising grew 10.6%.

https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm

- Anthropic: $895.4bn commitments/leases, mostly compute/infrastructure; $69.8bn due in the 12 months after 30 September 2026.

https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm#:~:text=an%20aggregate%20committed%20spend

- HackerTube: Karpathy’s project and write-up.

https://hackertube.org/

https://karpathy.bearblog.dev/hacker-tube/

Production: AI-generated intro voice and theme music.

Watch this episode on YouTube: https://youtu.be/irJLTdh3gPc

What happens to software subscriptions when you can build exactly what you need?

Dara and Matthew talk about AI making custom tools, websites and creative work more accessible, and what that means for the businesses that sell them. From replacing a WordPress site to the small-business “AI slop” argument, they weigh the freedom to build your own against the time, judgement and maintenance it still takes.

They also revisit their predictions, compare the latest AI assistants and model benchmarks, and discuss AI’s role in finding security holes that were already there. Plus the odd tangent and dangerously high fizz.

Recorded 23 October 2026.

Chapters:
00:00 Show intro
00:15 Back, predictions and preparation
07:21 The fizz reading
12:11 What do we want from an AI assistant?
17:29 Claude Duet and regional rollouts
19:37 New models and the benchmark problem
26:01 Google’s naming problem
32:29 Customising Claude Code
38:05 OpenAI models on Google Cloud
41:14 Dashboards, context and conflicting answers
45:01 AI and a week of security vulnerabilities
01:00:02 Software subscriptions and the Photoshop clone
01:05:10 Replacing WordPress and building the podcast site
01:11:38 Coding, images and content in one place
01:13:41 Small businesses and the AI slop argument
01:17:15 Stock images, freelancers and changing business models
01:23:48 Language learning and personalised tools
01:29:32 Predictions for the next episode
01:33:25 Outro

Links and resources:
Claude Mods: https://github.com/anthropics/claude-code/issues/91870
Feature overview and demos for customising Claude Code.
Cloudflare security-audit skill: usage and example prompts: https://github.com/cloudflare/security-audit-skill#usage
Cloudflare’s own instructions and prompts for auditing a codebase.
Grok Bot FAQ: https://docs.x.ai/grok-bot/faq
The personal assistant Matthew uses for household projects, reminders and email.
Claude Dashboards: https://claude.com/resources/articles/dashboards-and-motion
The interactive dashboard feature discussed in the episode.
Descript MCP: https://www.descript.com/mcp
The integration Dara uses for the podcast editing workflow.
OpenAI models on Google Cloud: https://cloud.google.com/blog/products/ai-machine-learning/google-cloud-brings-openai-models-to-customers
The announcement behind the cloud discussion.
Google Knowledge Catalog: https://cloud.google.com/blog/products/data-analytics/knowledge-catalog-self-learning-context-enrichment-for-ai-agents/
Shared business definitions and context for AI agents.
Cloudflare’s PHP2Shell protection guide: https://blog.cloudflare.com/php-vulnerabilities-waf-protection/
Background and protection guidance for the PHP vulnerabilities discussed.
Getty Images restructuring filing: https://www.sec.gov/Archives/edgar/data/1898496/000121390026112265/ea0309581-8k_getty.htm
The primary filing for the restructuring discussed in the episode.
It’s Only Logical website: https://onlylogicalpod.com/
Our AI-built podcast website, discussed in the episode.

Production: AI-generated intro voice and theme music.

Watch: https://youtu.be/AKXzS6CfLso

There’s always one more thing to try.

Dara and Matthew compare the promise of AI assistants with their own experiments in making them useful. They talk through agents, Paperclip, cloud environments, model routing and T3 Code, alongside subscription limits, context handovers and the temptation to keep tinkering.

When does one quick test become a fragile system, and when is it time to stop? Plus the odd tangent, an experiment graveyard and a fizz update.

Recorded 5 October 2026. Product availability, model names and limits reflect the conversation at recording.

Watch: https://youtu.be/aiq6JmURXrY
Chapters:

00:00 Show intro
00:15 Episode 2 and the name
01:11 Planning the episode and grinding axes
03:17 Dots, regional rollouts and the promised assistant
10:19 Coding AI and everyday knowledge work
15:57 Keeping up with AI news
17:23 Compaction, models and subscription limits
28:51 Clear answers and human oversight
36:33 Predicting the next AI releases
41:14 Always-on agents and the practical fallback
01:00:34 Multiple providers, CLIProxyAPI and T3 Code
01:05:46 Local models and routing
01:09:42 Context and handovers between providers
01:13:47 Tinkering and the experiment graveyard
01:17:26 The fizz reading and Ziggy
01:20:34 Outro

Resources discussed:

- Paperclip: https://github.com/paperclipai/paperclip
- OpenClaw: https://openclaw.ai/
- T3 Code: https://github.com/pingdotgg/t3code
- CLIProxyAPI: https://github.com/router-for-me/CLIProxyAPI
- OpenCode: https://opencode.ai/
- OpenRouter: https://openrouter.ai/

Production: AI-generated intro voice and theme music.

Dara Fitzgerald and Matthew Hooson launch It's Only Logical with their first reactions to OpenAI DevDay, recorded straight after the keynote on 29 September 2026.

From dots and GPT Space to new models, pricing and Codex in the cloud, we compare what looked useful, what still felt unfinished and where we want to experiment next. We also get into voice assistants, model routing, meeting transcription and the growing AI platform race — with a few air fryer tangents along the way.

These are our reactions on the night; availability and product details may have changed since recording.

In this episode:
- Dots, GPT Space and collaborative AI assistants
- GPT-6.1 Sol, pricing and usage limits
- The Decisions API and model routing
- Codex in the cloud, sign-in and apps
- Meeting transcription and local voice tools
- Google, AI hardware and privacy

Chapters
00:00 Intro
00:15 A new podcast (and an air fryer tangent)
04:14 OpenAI DevDay: first reactions
08:51 Dots: always-on AI assistants
13:33 Voice assistants and “let me check”
16:15 GPT Space and collaborative agents
27:20 GPT-6.1 Sol, the model race and pricing
37:26 The Decisions API and model routing
43:13 Codex in the cloud
47:04 Sign in with ChatGPT and apps
50:38 Meetings, Wispr Flow and local transcription
55:00 OpenAI Marketplace
57:16 Google and the AI arms race
1:02:00 AI hardware, privacy and AR
1:07:03 Wrapping up

Mentioned in this episode:
OpenAI DevDay recap: https://openai.com/index/devday-2026-recap/
Wispr Flow: https://wisprflow.ai/

Hosted by Dara Fitzgerald and Matthew Hooson.
Contact: [email protected]
Watch on YouTube: https://www.youtube.com/@onlylogicalpod

Production note: the intro narration and theme music were generated using AI. The episode conversation is Dara and Matthew's recorded discussion.