Twitxy Saved

Joshua Ness

@jcness

AI strategy and governance for private capital, financial services, and legal. AI Specialist at @FulbrightPrgrm. MD at @StartupGrindNYC.
Brooklyn, NY joshuaness.com Joined July 2008
1K Following 1.9K Followers 11K Posts

Latest posts 5

Joshua Ness @jcness
Did you know there were recently 4 AI supply-chain security incidents in 50 days? They hit OpenAI, Anthropic, and Meta-adjacent infrastructure.... things like elease pipelines, CI runners, and dependency hooks.

It got almost zero coverage, which is unfortunate.

The attack surface for AI systems is turning out to be the software factory around the model... not necessarily the model, itself.

Red teams test for things like jailbreaks and system card violations, but they don't typically test the publish gates, package managers, and CI pipelines that ship the models into production.

While I wish this were a coincidence, four incidents in 50 days across three major labs is systematic. When something breaks publicly (and it will soon) this will suddenly get all the coverage it deserves.

Check out the @VentureBeat article for a bigger breakdown.
venturebeat.com/security/suppl…
venturebeat.com AI supply-chain attacks bypass model red teams Four supply-chain attacks hit OpenAI, Anthropic, and Meta in 50 days — none inside the model. A 7-row matrix maps what AI vendor questionnaires are missing.
Joshua Ness @jcness
Notion just opened its workspace to outside agents. Anthropic just locked Claude's agents inside your own infrastructure.

Same week, opposite architectures. One of them is wrong about where enterprise trust lives.

@NotionHQ's bet: the platform is the trusted intermediary. Any agent, anywhere, can plug in.

@AnthropicAI's bet: enterprise buyers need control at the perimeter. Your data never leaves your environment.

Whichever architecture wins will eventually set the default for how AI agents integrate with corporate workflows through 2028. It's governance philosophy masquerading as a product decision.
Joshua Ness @jcness
This week, Google I/O announced Gemini as the agentic runtime embedded in Search, Workspace, Chrome, and hardware.

On the same day, Anthropic grabbed @StainlessAPI (SDK layer) and announced @karpathy will lead AI-powered training automation.

Neither company is competing purely on model benchmarks anymore. Both are racing to own the full stack: training pipeline, developer tooling, product integration, and hardware (Nvidia Vera arriving at both companies simultaneously).

Frontier model quality is becoming commoditized. The companies that win this phase will be the ones who have the most locked-in infrastructure.
Joshua Ness @jcness
Two different AI papers this week made the same point from opposite directions.

@thinkymachines says the real frontier is real-time interactivity over static output quality. Josh and Palmer at @AICollectiveCo say it's recoverability, no matter if failures are visible and bounded.

Both are right. Neither shows up on a benchmark. The field has been measuring the wrong things. Deployment is forcing the issue.

The practical implication: If you're evaluating AI tools using vendor benchmarks or public leaderboards, you're selecting for performance in conditions that don't exist in production.

The model you pick because it aced MMLU might be the one that wipes a folder without telling you, or breaks when your users don't communicate in tidy, complete prompts.

For anyone making AI purchasing or deployment decisions, the honest criteria aren't "what's the benchmark score?" They're: "What does this model do when it's wrong? Does it tell me? Can I recover from it?" Almost no vendor answers those questions directly, because almost no benchmark has required them to.