I write this as an Open Letter with feedback about the Sovereign AI innovation process. It is addressed to the Sovereign AI team and the civil servants who own the challenges — experts in their respective fields — potential future coworkers, colleagues, project partners.
Please share among your colleagues 🙏
Feedback and ideas are a team sport 🙏
This open letter offers constructive feedback directly to the challenge owners and civil servants designing the pipeline— because unlocking genuine frontier innovation requires removing friction before the October deadline.
I often communicate through images, as they say image is worth 1000 words, that's why attaching my 1st feedback as an image:
"Wildcard" or "surprise us" category, as opposed to strict definitions.
Another alternative is broad "catch all" category such as "AI in public service".
This will enable anyone who uses public services to come up with ideas.
On the same note: information gathering from the people who are users of the systems.
Optionally: do not require name to encourage more unfiltered answers.
These days "The hottest new programming language is English" (tweet by Andrej Karpathy) and I genuinely believe that AI literacy classes would be super practical across the entire society. To create a 1st draft of a software takes 1 prompt and 5-10 minutes of patience, not exaggerating.
Improve the challenge descriptions.
Currently there are narrow criteria with very specific wording.
(screenshot from latest guidelines with the deadline 16 October 2026, another version with the deadline 7 May 2026 is here, change in document indicates this is a living process and the feedback coming from this Open Letter has a chance of landing)
Challenge 2 wording feedback: "Platform" here means a tank, ship, aircraft or drone. In software it means the product a startup is building. "Edge" appears twice in two sentences with two meanings — "competitive edge" (advantage) and "edge AI deployment" (running on the device in the field) — while in AI the word usually means frontier, as in frontier models.
Challenge 3 wording feedback: "Drive efficiency and effectiveness in public compute infrastructure" is a challenge. Anyone with an idea for making public compute cheaper could answer it. The rest of the sentence takes that away. "To help deliver expansion of the AI Research Resource" limits the scope. "Through trialling and deploying novel hardware or software supporting compute efficiency and heterogeneous systems development" limits the method. Little place for innovation left.
Reduce gatekeeping layers.
I personally identified a few. Then asked AI to fact check and found even more. It is so complicated it sounds funny :)
Across an multiple stage application funnel, applicants face multiple redundant administrative hurdles before building anything. I suggest using 2 weeks to build prototypes, not to assess initial expression of interest letters.
Both stages 6 and 7 operate on a threshold basis, effectively pass or fail. The actual decision is made by committee in stage 8, without participation of the challenge sponsors.
Why is it this complicated?
From an applicant's perspective, it appears that multiple governance models may have been layered together: grant administration requirements combined with procurement exemptions and multi-stage committee approvals. Regardless of the exact internal structure, each layer introduces additional friction without removing earlier gates.
Loads of money to spend. Minimum check size. Operational complexity. Communication overhead. The first umber that came to mind: 100 projects. I like 100 projects as "reasonabe" baseline, within the Dunbar's number over time, as capacity grows it could become 100 per challenge.
Communication as a feature, not an afterthought.
Three quarterly updates published on the same date (link to Substack) appears odd. UPDATE: #1 #2 #3 on the personal blog of the James Wise, who chairs the fund and is a partner at Balderton Capital.
Online briefing that ran out of capacity: "This event is currently full"
Lack of responsiveness on Twitter:
Our challenge based R&D Procurement Scheme is LIVE with the first challenges coming from @ARIA_research, @DefenceHQ, @NCSC and @DHSCgovuk .
We’re creating new
Now reaching out to the James Wise on LinkedIn and Twitter.
Scalability beyond the UK should be used as evaluation criterion.
Some things like birth certificates, death certificates, marriages, land registry, vehicle registry, driving licences, voting are similar across the entire world. What started as grant award can become the global digital infrastructure for all of humanity.
Big ambitious idea to attract the best talent from the entire world, potentially leading to UK employment visa.
Auditable in real time. Possibly some exceptions for "skunkworks" secret military budget, otherwise public by default.
People participating in decisions weekly, not once every few years. Prototype built at PlanetaryCouncil.org/vote some questions are universal for the world, some questions are UK specific. It starts with proportional representation that is widely supported across the political spectrum.
Surely something like this should exist (maybe it already exists). Transparent and unbiased, not provided by some think tank with questionable funding and agendas.
Currently the House of Commons seems to me like a museum of the old system:
we should be able to vote sensibly like everyone else would think we would do
Just like new design of House of Commons, this idea is also beyond the scope of AI but kind of obvious from the root cause analysys perspesctive.
Please stay in touch. And please share this Open Leter among your colleagues.
If you are on the Sovereign AI team or lead one of the challenge areas, let's discuss how these recommendations can be implemented in the current funding cycle.
My email is public: marsXRobertson@gmail.com
My ☎️ is pretty public too: book a time via Calendly
Open for work. Working with smart people on impactful projects = 100% aligned. I would love to use my skills towards planetary regeneration, starting from UK 🇬🇧
Open for business.
Open by default.
In my previous career I was pretty good with IT, top 3% on StackOverflow.
Now I specialise in using AI for images, music, videos, code and deep philosophical debates with frontier models. Here is my latest work, I hope it will inspire you. Sound on.