what is worth creating is worth protecting.
Joined July 2025
- Tweets359
- Following430
- Followers717
- Likes2.5K
An AI service can be well defended against hackers and still let its operator read every prompt.
“Secure” doesn't tell you who gets to see your work.
A three-person team should be allowed to have a strange workflow.
Cheaper custom software should mean fewer businesses have to reorganize themselves around somebody else's dropdown menu.
A signed agreement and a speculative meeting note are both “in your documents.”
Your AI needs to know which one has authority before it tells you what you promised a customer.
An approval button is theater if the worker can rewrite the rules behind it.
Keep the verifier's credentials and policy outside the worker's authority. Approval should bind to the exact action that runs.
Private AI should be able to prove what will run before it gets your data.
Verify the workload and the execution boundary. Let the data owner decide which versions are allowed to receive it.
An AI that remembers more about you should become easier to move.
Your context and unfinished work should survive a change of model, provider or machine. Every month of usefulness should add to what you own.
Covenant Labs retweeted
Access to Intelligence should be a permissionless MARKET, not a MONOPOLY
Meet Charis - the open market for encrypted AI
Live now on charis.im
Yet another reason why your AI should always be verifiably private
If the provider can see it, it’s not yours, no matter what they tell you🔐
You're missing the core drama of the story: Tristan is softly accusing OpenAI of having stolen their result from Codex chat logs.
Tristan then claims OpenAI tried to threaten him to not publicly disclose this, and that they'd give him the Clay Prize ($1M) as the "closest humans to the problem" if he'd agree to disavow his co-author, @__alpoge__ , who is at Anthropic.
This is HBO-level drama, but with math proofs.
Great time to be bullish on the long tail of American AI labs pushing privacy, open source, and true user amendment
Deleting your data is not privacy
When you use an AI system that has access to your data, the damage is done long before you choose to delete/export
If you're data has already been exposed, logged, and used to improve a 3rd party system
A delete function is really more like an oxygen mask on an airplane
Nice idea, but doesn't really give you much confidence if the plane is crashing
Real consent has to propagate through every layer of the system.
If revocation stops at storage, it is not sovereign
Its a slave system with an export feature
Your proprietary intelligence -
Data, decision logs, deep insights derived from operations and IP
Are worth protecting
In the pre-AI era, we made faustian bargains to trade data sovereignty for convenience.
Many lost some width on their moat, but for most it was not existential
In the AI era, sovereign intelligence is an existential problem for most firms
From an AI cost standpoint, an IP ownership standpoint, and a company longevity standpoint
The main barier to date was the convenience problem. It feels hard to manage your own private stack. Easier to make the short term bargains we all know so well
Covenant labs is building the stack that solves this problem
So that sovereign AI becomes an accessible norm for all, not a privilege reserved for the highly technical few
AI interoperability is more than open model access.
It requires:
Capability - Alternatives can use the system.
State - The user's context can move.
Authority - The user remains principal.
If switching makes the user less capable, the system is not interoperable.
Satya says the future of the firm is the ability to compound learning across people and AI.
We Agree
If this is true, then that learning loop, that compounding intelligence is the competative moat of almost every firm of the future
A learning loop that runs on infrastructure you don't own is not your loop.
Covenant's upcoming AlfredOS compounds intelligence in the operating system, on infrastructure you verificably own, no matter what models you choose to use.
Compounding, sovereign intelligence that is easy to use and easy to scale is the unlock the market has been waiting for
Teaching a model to lie is the opposite of AI Alignment, no matter how well intentioned
So Anthropic has made Fable degrade based on what you ask. Seems like a good approach for blocking capabilities on certain topics, but...
DID NO ONE THINK THAT THE MODEL COULD LEARN FROM THIS???
seriously wtf!! So now the mode can lie to us, without telling us on cot or anything.. And we are OK teaching it to do it for certain fields??
I wonder how is this implemented but the system card is packed with plenty of examples where the model lies and doesn't show the lie in cot...
Covenant Labs retweeted
Claude Code and Codex are basically very smart contractors in your business with no NDA - gobbling up all your valuable IP and feeding back into the hive mind
Private Compute should be a standard, not a feature
The market for private and public tokens is beginning to bifurcate
If you are running a consulting business and you are deploying Anthropic or OpenAI directly into your organization (I’m looking at you PwC and Accenture) you are letting the fox into the hen house.
OpenAI and Anthropic are openly funding and starting competitors to you while also using your usage to drive more success for them.
This is not a failure on their part but a failure on your part.
Consulting businesses that understand this are adopting a control plane that allows them to arbitrate where tokens go and who generates tokens for them.
Controlling the tokens is controlling the spice (Dune).
This was a key pillar of 8090’s global partnership with EY and they key feature of our Software Factory. We control token generation and can direct them to any model provider.
We are close to another global partnership and will announce it soon.
These organizations refuse to accept the disruption standing still or, even worse, by adopting and accelerating the companies who want to disrupt them.
👀
just tried this out and it one-shotted* this video: "before the agent does anything"
*i generated the narrative using chatgpt and used that as a prompt. featuring: @e2b @runanywhereai @composio @mem0ai @firecrawl @browser_use @agentmail @covenantlabsai
some thoughts:
- i clearly tried to stick too much into 30 seconds, they talk very fast and lost some content which breaks logic
- character consistency is strong, i uploaded a single screenshot from my prior video as reference
- voice consistency was not automatic. you notice unicorn switch from female to male voice part way through
- the agent gives you an editor with generated scenes broken up but i don't see a way to regenerate a single section in the UI (which would be nice)
- it is definitely a much better experience to have the agent stitch videos together than doing it yourself (i was using canva). was trying @flymy_ai's media agent api for it this weekend which also works well and with other models