Wayne

Building and experimenting with AI Agents

[Wayne @wayneb]

@thecsguy Elon needs to throw some more of that compute at it

[Wayne @wayneb]

PS: Here is my Podcast Clipper for X template: https://t.co/qDOU3Q10cD

[Wayne @wayneb]

5 Awesome Grok Bot updates this week

- Pages: reports and dashboards you can edit together
- Video: describe it or drop an image, Bot cuts the clip
- Vault: save logins, cards and contacts once
- 1Password: approve a login, Bot signs in without seeing the password
- Shared Bots: one team Bot for CRM, quotes, deal terms

Highly recommend watching @mattyp's walkthrough of this wave (88K view already - nice one Matt!)

I’ve been using it daily since launch. The pace is getting silly

Which of these are you actually using? 🤔

[Wayne @wayneb]

Fable 5.5?

A mystery model called "Obsidian 3" just turned up in Cognition's FrontierCode 2.0 data

and it scored 46.8%, just ahead of Opus 5.5 at 45.5%

It runs in the Claude Code harness, which only Claude models use.

Nothing confirmed yet, think its Fable 5.5? 🤔

[Wayne @wayneb]

@firesidealpha Love this!

Give kids a real stake in the future instead of a board deciding for them

Elon, Dell, Gerstner… more of that please 💪

[Wayne @wayneb]

Dario Amodei just said the quiet part about AI pricing

Claude telling you to restart your Mac: worth a dollar.

Claude telling a pharma team where to move a ring on a molecule: tens of millions

Same tokens. Wildly different bills if he gets his way

Pay-for-results sounds fair until someone else prices your outcome

Would you take that deal?

[Wayne @wayneb]

Grok n Bot - X just started paying referrals in usage resets

Bring someone onto Premium. When they actually pay, you get a full usage refresh

How it works:
- Share your personal referral link
- They get 14 days of Premium free
- First payment clears, you earn one usage reset (7-day hold)
- Reset restores included usage for the current period
- Rewards last 6 months
- No cap right now

US only. Web checkout. Premium, Super or Ultra.
App Store / Play Store plans don't count

I burn through Grok Bot limits most weeks and this is the first referral program that pays in the thing I actually run out of

Would you prefer a reset or $$? 🤔

[Wayne @wayneb]

@MaximLevoshin This is a topic I always think about because the UK education system has so many gaps

[Wayne @wayneb]

90% of Claude Code is written by Claude, here's how they stop it shipping bugs

50+ review rounds on one pull request.
Two AI agents arguing all night until a human stepped in and stopped it 😂

That's from the Claude Code team's write-up on how they review code when Claude writes about 90% of it.

What actually works for them:
1. Claude has to show proof it works (tests, build, screenshot), not just say "done"
2. A fresh agent reviews the diff before any human sees it
3. Billing, CI and other risky stuff always gets a human
4. Bot approvals never count as human ones
5. After 3 rounds the agent has to rethink the approach

Merged PRs more than doubled June to September. Reverts fell 40%+

Writing code got super cheap but now the job is deciding what to check

[Wayne @wayneb]

Ummm, Astra-level computer use at 1/10th the price?

GPT-6.1 Terra matches GPT-6 Astra on Agents' Last Exam at 1/10th the price per token.

And OpenAI says it's over 3x faster too

$1 in / $5 out per million tokens.
That's literally half the old Terra

The number I'd actually circle is this from Artificial Analysis. When it doesn't know the answer:

- it now guesses 38% of the time
- vs the last Terra which guessed 88%

For an agent clicking around your apps, a wrong click costs you more than a slow one

It's not OpenAI's smartest model though. Sol and Astra still win on hard coding and research.

That's kind of the point though...

The race has moved from smartest model to cheapest worker you can trust

Wen Grok 4.8? 👀

[Wayne @wayneb]

🚨$40B ("B" for Big ones!)
SpaceXSI just locked the largest GPU financing ever

It's a master finance lease for Nvidia Vera Rubin chips
Closing targeted before year end
Gear starts arriving H1 2027

CFO Bret Johnsen says:

"We're seeing less than 1 year payback on SpaceXSI infrastructure".

Honestly, I think that's wild for metal this expensive

SpaceXSI are leasing the picks and shovels at scale so they can keep selling compute to labs like Anthropic

Debt markets are super tight for most AI shops yet SpaceXSI still got $40B. How? Why?

-SpaceX has an investment grade credit rating
-Starlink cash

How long before every serious AI lab needs a financing partner this size just to stay in the game? 🤔

[Wayne @wayneb]

This is how I think the Anthropic IPO plays out through 2040. WDYT? 🤔

[Wayne @wayneb]

@StockSavvyShay I think this is what the general performance will look like

[Wayne @wayneb]

@theo I thought Grok updates were supposed to be every week or every other week?

[Wayne @wayneb]

Anthropic's IPO filing just handed SpaceXSI a $120B compute cheque

Up to about $3.1B a month (once the GPUs are live), so approx. $37B a year through 2029

Capacity is meant to be fully delivered by December.

Same terms apply, after early 2027, either side can walk with 90 days' notice (more an insistence from the SpaceXSI side rather than Anthropic as Elon pointed out).

Best way to think about it is like a lease on the whole datacenter, not a cloud tab you can quiet-quit

This is another almost $40B of revenue for SpaceXSI and from a user perspective, I really like this partnership

If you had to pick one, would you rather own the lab burning $37B a year or the one collecting it? 🤔

[Wayne @wayneb]

@StockSavvyShay The interesting part is who's on the other side of that bill.

SpaceX and Google get paid either way

[Wayne @wayneb]

@elonmusk @bot Bring phone calls to the UK and Europe early. That's the feature I'd use most this side of the pond.

Also, a clear usage meter per routine. I'd run way more scheduled jobs if I knew what each one costs me a week

[Wayne @wayneb]

@thecsguy @Phil_Holland This is great man.

Here is my Podcast Clipper for X:
https://t.co/qDOU3Q10cD

[Wayne @wayneb]

@mattyp @bot I was thinking that as sharing bot templates scales, it would be great to know how many tokens/usage it eats

[Wayne @wayneb]

Before you switch to Kimi K3.1 Pro because of your feed

Open weights don't land until Nov 4

And Derya Unutmaz says in his testing it isn't even cheaper per task than Opus 5.5 on subs.

Test it yourself first