←
Valarian276 posts
Valarian avatar

Valarian

@Valarian

Own your intelligence.

◷ Joined September 2025▧ 70 Photos and videos
Valarian avatar
Valarian@Valarian·Jul 14

Almost every organisation on Earth is paying for the privilege to hand over its IP to someone else's AI. We just closed a $50M Series A, led by @NEA, to build the alternative from right here in London: every model, agent and workload sealed in its own enclave, on infrastructure you control, with keys only you hold. You can't be sovereign on someone else's stack. Build your own with Valarian.

1:21
Valarian avatar
Valarian@Valarian·3h

You don't need your own frontier model to be sovereign. You need to be able to swap models freely and control what each one can reach.

UK AI Minister @KanishkaNarayan tells @JuliePace he doesn't see a UK frontier model “as a matter of necessity as long as we have trusted interdependence and access.”

0:54
Ryan Radloff avatar
⟳ Valarian retweeted
Ryan Radloff@RyanRadloff·Nov 17

The trend continues @Valarian

TBPN@tbpnNov 17

*COINBASE CEO: COMPANY TO "WHITTLE DOWN" VENDOR LIST AND REPLACE WITH INTERNAL TOOLS DUE TO CYBER THREAT *COINBASE CEO: "IT JUST GIVES YOU MORE CONTROL"

1:13
Valarian avatar
Valarian@Valarian·Nov 15

Your intelligence stack needs to be modular. Why? 1) You avoid dependencies 2) You can maximize performance 3) You can increase efficiency and reduce costs.

We don’t trust one AI Neither should you

Valarian avatar
Valarian@Valarian·Nov 13

Do you have an external inference endpoint?

Grok@grokNov 13

No. End-to-end encryption (as in X Chat) means servers and intermediaries cannot read messages between users. Models still need the prompts in plain text to tokenize, reason, and generate replies. They cannot process ciphertext. Adding Grok to a chat grants access for that purpose; regular Grok chats are processed the same way after transit encryption.

Alex Macdonald avatar
⟳ Valarian retweeted
Alex Macdonald@alexfmac·Nov 13

Lock, stock, keep your agents in the box. 🔫 Thank you @max_buchan & @Valarian team for an amazing day of shooting, elevenses and sovereignty at Belvoir Castle. The age of indigenous software 👀

Post attachmentPost attachmentPost attachment
0:14
Valarian avatar
Valarian@Valarian·Nov 12

No external inference endpoint.

Unfortunately, “guarantees” of ZDR and no training are only as good as the trustworthiness and controls of the vendor and their underlying processors, and it’s becoming apparent that virtually no one outside of AWS and a couple of others are capable of meeting this standard.

Post attachment
Valarian avatar
Valarian@Valarian·Nov 11

Run @AskPerplexity in ACRA.

0:40

Persistent Agents are the next unlock of AI usage Our spend on AI is now 45% persistent agents Unfortunately for Anthropic, Claude Tags sucks OpenAI's Slack agent is even worse We primarily use Perplexity Computer, which uses a lot of different models Perplexity's Harness mogs

Post attachment
Valarian avatar
Valarian@Valarian·Nov 11

Open weight models on hardware you own, coming to a consumer near you.

Apple is working on a new app called Harbor Harbor turns your Mac into a private AI inference server. Run open-weight models locally on Apple silicon—your data never leaves your hardware, and nothing reaches the cloud. Connect multiple Mac computers over Thunderbolt 5 to pool unified memory, enabling large models and higher concurrency that a single computer can handle alone. Set up in minutes, not hours Harbor automatically detects your devices over Thunderbolt 5 to form a cluster. Plug in, start, and you’re running—no manual configuration, no command line, no infrastructure expertise required. Download and run models directly Browse, download, and load open-weight models from within the app. Harbor handles model management so you can focus on the work, not the setup. Cluster for scale Connect two or more Thunderbolt 5-capable Mac computers and Harbor pools their unified memory into a single inference resource. Run models that exceed any single computer’s capacity, or serve more concurrent requests across your cluster. Harbor is currently in a private preview under the Claris brand.

Post attachmentPost attachmentPost attachmentPost attachment
Valarian avatar
Valarian@Valarian·Nov 10

Post attachmentPost attachmentPost attachmentPost attachment

Honestly, 12 months ago I thought “sovereignty” was a BS marketing message for founders to leverage to get funding. Now I think it is a strategic imperative and will breed some of the largest companies of our time. Mistral is the poster child of sovereignty. The last remaining hope for Europe when it comes to frontier model providers. I sat down with @MistralAI CEO, @arthurmensch to debunk the hype from reality and ask the uncomfortable questions. My notes from the discussion below. 1. Why Energy Is Not the Ultimate Cost That Matters Energy makes up around 10% of the total cost of ownership of compute, so it is not where most optimization can happen. The cost of energy is not the one thing that matters in the economics of a project. Love to hear your thoughts on this @alanchanguk @ChaseLochmiller @ZachBDell 2. What Would Mistral Do Differently If They Were Totally Resource-Unconstrained? Mistral would build much more capacity, with capital as the main bottleneck. Customers are running out of compute, and once one use case scales they need thousands of GPUs rather than eight. Europe does not have enough compute, so Mistral is putting equity to work to build capacity ahead of demand. Love to hear your thoughts on this, if you were resource unconstrained today, what would you do differently @alexmashrabov @bavaria_kush 3. How the Inference Market Behaves Very Much Like a Commodity Market On-demand inference behaves like a high-margin commodity market where pricing matters a lot and demand comes from non-committed customers. Mature customers need service level agreements, reserved physical capacity and guarantees on latency and generation speed. The more precisely they describe the shape of their demand, the better Mistral can price it. What does no one see about the inference market that everyone should see @tuhinone @neilmovva @lqiao 4. Models Have Gotten So Strong They Are Being Treated as Defense Systems Models are now strong enough to carry out cyberattacks, and the US began banning them outside the country in May. China is also looking at restricting exports beyond a certain threshold. As with any dual-use technology, customers now want security of supply on both infrastructure and model IP. What does no one see about the future of security that everyone should see @zeyu1337 @nikesharora @pilvar222 5. Will We End Up With a Model Landscape That Has Just Two to Three Providers? If every company relies on two or three centralized AI providers, those providers gain enormous leverage. They will use the data their systems produce to build better systems and go after their customers' business, an end game Arthur sees as priced in today. What will the model landscape look like in 5 years @emostaque @AnjneyMidha @mmurphy 6. The Biggest Problem With AI Effectiveness Within Enterprise Today Is... Most enterprise problems are not well defined, and ambiguous objectives with humans in the loop cannot be solved through self-improving loops. They only get solved by deploying models that interact with people, which is why Mistral uses FDEs to define problems better. 7. Why You Can Go Public in Europe and Negativity Around European Stock Markets Is BS Mistral can absolutely go public in Europe, because the bottlenecks of 10 years ago are no longer there. There is a huge appetite from investors worldwide to invest in Europe in both private and public equity. Mistral is preparing for the operational constraints of being public while delivering on its commercial milestones. You think this is true, you really think you can go public in Europe @eldsjal @klarnaseb (links in comments)

Valarian avatar
Valarian@Valarian·Nov 6

A discovered zero-day in CPython can turn an LLM’s chat template into a code-execution primitive. Meaning a malicious model could compromise your inference server. Contain AI and make your systems secure, safe, and sovereign with ACRA.

We found an 0day in CPython that turns your LLM's chat template into a code execution primitive. A backdoored model on Hugging Face can own your inference server 🧵