Builder, starter, tinkerer, consultant. ⛑️ ResponseHub (AI automation for security questionnaires) 👨🏻✈️ Applied AI consultant 👷 Product dev
London
Joined December 2008
- Tweets5.9K
- Following2.2K
- Followers1.3K
- Likes5.6K
Never thought I'd see PHP on the front page of BBC news. Thoughts and prayers to all the PHP shops out there.
This British mindset is one of the main reasons why I left the UK 3 years ago, but so much has changed since then. I’m convinced that you really just have to intentionally curate the community of blacksheep misfits with the delusional optimistic mentality of the Americans and you can move mountains.
Gotta stop blaming everything and everyone else and start taking accountability to make shit happen.
Just moved back and working on this rn, stay tuned…
For us old-hand web devs, this keynote will go down as the moment we closed the old chapter of our industry and started a new one.
IMO @dhh nails the mindset we need to take up:
what a privilege it is to straddle the era of hand writing code and the era of steering intelligence to write code for us.
there's grief in saying goodbye to the old ways of doing things, clinging on to that is futile. There are two ways to move forward:
1. Change careers
2. Get unreasonably excited about what we can make in this new world.
I'm choosing option 2.
My Rails World keynote will be streaming in a few hours on YouTube. I dare say this one might ruffle a few feathers, but it also just needs to be said. youtube.com/live/V9SxpJpHuus…
Ii switched ResponseHub from credit based to seat based + unlimited usage.
I was finding that credit based pricing discouraged product usage. Users apply a $ value to each credit and then question the value of each output. Sometimes this can work but not always.
Pricing is fucking hard for AI native apps.
per seat is dead, the seats are agents now. per ticket rewards volume you should have prevented. per resolution punishes you for having a good product.
we're settling in at no seat structure, just credits. put the whole team on it for free, pay for what the agents actually do.
might redo it in 3 months. lets see
The watermark is… “load bearing”
🚨 HUGE NEWS: Claude now embeds an invisible watermark into every piece of text it generates.
Anthropic just documented how it works. Two marks, both machine-readable:
> Text: an imperceptible watermark woven into the words themselves. You can’t see it, and it doesn’t change meaning, quality, or readability.
> Files (.svg, .png, .jpg): signed provenance metadata on the C2PA open standard, so you can tell if a file’s been tampered with.
The watermark is applied at the model level.
That means it shows up no matter where the text comes from: the API, Claude, Claude Code, Cowork, Claude Tag, and even when a supported model runs through AWS, Google Cloud, or Microsoft Foundry.
Models launched on or after August 2, 2026 mark from day one. Older models are getting it during a transition period.
Every sentence Claude writes for you now carries a signature you’ll never see.
The price of GPT-4 level intelligence has fallen 13x in four months and 1000x over 18 months, driven by GPT-5.6 Sol recursively optimizing its own serving infrastructure.
latent.space/p/ainews-gpt-56…
TIL footballers have their own evangelical movement / cult
England and France gave everything in the World Cup third place play-off, but after the final whistle, players from both teams came together in prayer.
The players praying together were: Dayot Upamecano, Bukayo Saka, Maxence Lacroix, Eberechi Eze, Marc Guéhi, Jean-Philippe Mateta, and Trevoh Chalobah. ❤️
If you want to catch up on all the latest Kimi news, spend a few mins browsing Lollygag. All the AI news from the outlets that matter, summaried.
lollygag.news/
Any good ideas for Fablemaxxing before July 19th?
Currently on my list is:
- one shot SAML SSO implementation
- one shot SCIM user provisioning
- audit and fix all Turbo / Hotwire / Stimulus stuff
- performance audit
- security audit
- one shot implementation of new marketing site design
Fable gave me some more:
- Audit every background job for what happens when it fails, retries, or runs twice.
- Audit the AI pipeline end to end: every call into Bedrock (Claude/Nova), Mistral, Textract, and the reranker. This is a reliability, cost, and safety review.
- Audit the app's query patterns for the performance failures that degrade gradually: N+1s, missing indexes, unbounded result sets, and per-record work in loops.
- Audit every AASM state machine for states a record can get stuck in and transitions that can misfire.
- Audit acts_as_tenant usage
X timeline definitely improved. Less doomerism spam, more accounts I follow. More interesting content. 👏👏👏
Millwright, an open source cloud based software factory, now supports naming specific models for specific tasks so you can queue up your Fable tasks.
github.com/njcameron/Millwri…
Every time I start a new project I feel I should be using TS / Next / vercel / ingest / etc
And then I go and choose Rails.
This week we're rolling out a huge change to how we charge for ResponseHub.
When we started, I took inspiration from products like Fin, the customer service agent, which did something revolutionary: it only charged for successful outcomes. I wanted to do something similar for ResponseHub and only charge for successful answers to security questionnaire questions. It felt like it would be win-win: customers only pay when they get a result.
The reality was more complicated. What if I regenerate an answer? What if the answer comes straight from the knowledge base? But most importantly, it was disincentivising product usage. I was hearing from customers that they wanted to onboard users from Sales and Customer Success teams so they could get answers to buyer and customer questions about security, compliance, legal and product.
So this week we're making three big changes:
1. Every plan now includes unlimited usage.
2. We've introduced a new role, "Viewer", who can ask questions and view content. Every plan includes unlimited Viewer seats.
3. We're now charging by "Editor" seats, the folks who manage the knowledge base and work on RFPs and security questionnaires.
The most exciting thing about this change is that it moves ResponseHub away from being a tool for security questionnaires and towards being a company brain that your whole organisation can use to answer buyer questions about security, compliance and product.
Building lollygag.news was the perfect opportunity to experiment with some of the new, near-frontier models. Lollygag summarises around 190 articles from 50 different sources each week. Since this is (currently) a non-revenue-generating project, I didn't want to end up with a huge token bill.
To compare options, I set up an eval in Promptfoo across several models (Sonnet, Haiku, DeepSeek, Gemma, GPT-OSS), and used Opus 4.8 as the LLM judge to review the summaries for accuracy and style (i.e. avoiding sounding like an LLM).
Here's the interesting thing: DeepSeek's performance was as good as Sonnet, but 39x cheaper.
It's totally exploded my assumption that using (near-)frontier intelligence in a product adds significant variable cost, and therefore impacts how you charge (usage-based vs fixed-fee) and how you go to market (freemium vs free trial).
Another way to think about it: Lollygag's hosting costs are $5.50 per month. If it relied on Sonnet, inference cost would be 1.3x the server cost; with DeepSeek it's 0.03x.
If you're a tokenmaxxing org spending six or seven figures on Anthropic tokens, I don't think you can afford to ignore these new models.
