Open algorithms, closed weights, and the arithmetic of AI tooling

00:00:04 Good morning, Friday the 15th. I'll start with X's decision to open-source its recommendation algorithm, then work through what that means. The gap between what's public and what's not is where this one gets interesting. X published the full code for its For You feed's ranking algorithm on GitHub under the xai-org organization.

00:00:26 The code acts as scaffolding — it loads features, calls models, and handles the ranking pipeline. It leaves out the ranking model itself. That's a trained model built with the same transformer architecture that powers Grok's Phoenix, and it stays proprietary. Daniel Meacham put it clearly in a reply: the code is open, the ranking model it calls isn't, and that's the part that actually decides your feed.

00:00:53 The transparency is meaningful but partial. You can audit the structure, the features that get pulled, and the way the pipeline works. But you can't look at the weights that make the actual ranking decisions. That's worth tracking. If you've ever wondered why certain kinds of posts show up more than others — and I have, on plenty of days — having the code around the model means you can see what signals X is measuring.

00:01:21 That's different from the black-box pattern most platforms operate under, and I haven't seen another major social platform do anything like it. The gap is wide. It's just incomplete. The question going forward is whether X keeps updating the code publicly every few weeks, as they've promised.

00:01:41 The one-off open-source event is a signal. The ongoing transparency is where the actual signal lies.

00:01:47 Next, a different kind of infrastructure story — one about how we build with AI today. Cloudflare posted a benchmark comparing SDK-based agent coding against MCP-based approaches. They tested it on their own GraphQL API. The results: using the SDK directly took one step and 15 thousand tokens.

00:02:09 Running the same task through an MCP server took four steps and 158 thousand tokens. Same output, 8.4 times the token cost. Yoni Braslaver and the team ran this on Monday's API update. It's a concrete, reproducible comparison, and the result favors the simpler approach: direct SDK bindings over the tool-search-plus-dispatch pattern that MCP was supposed to standardize.

00:02:37 The reason this matters is that the MCP ecosystem is still building out and the tool-search approach has real appeal. You can plug any MCP-compatible server into any agent without writing custom bindings. That's the trade-off: convenience and interoperability on one side, token cost and latency on the other.

00:03:01 For code generation tasks, Cloudflare's data says the direct route wins clearly. Tim De Pauw pushed back in the thread, suggesting this might be more of an argument for tool search than against MCP entirely. He was expecting a benchmark between agents using MCP versus agents calling the same tools through dynamically generated adapter SDKs.

00:03:26 That's a fair point — the comparison isn't MCP versus nothing. It's MCP versus a well-maintained SDK. Even if tool search narrows the gap, the order of magnitude difference here is significant. Anyone building an agent that needs to talk to APIs should keep the Cloudflare result in mind.

00:03:48 SDK-first, MCP-as-extension. Not the other way around.

00:03:52 On the academic side, arXiv just announced a new policy worth noting. Thomas Dietterich, the arXiv moderator for cs.LG, clarified penalties for papers containing unchecked LLM-generated output. The penalty is a one-year ban from arXiv, and after that, future submissions must first be accepted at a reputable peer-reviewed venue.

00:04:16 The policy targets incontrovertible evidence — hallucinated references, LLM meta-comments left in the final draft, and that kind of thing. Dietterich's wording is specific: if a submission contains this kind of evidence, it means the authors didn't check the results of LLM generation, and that's a breach of the Code of Conduct.

00:04:41 The Machine Learning subreddit reaction was mostly positive — one comment got 181 upvotes arguing for even longer bans of all co-authors, which is probably too broad. But the direction is clear. arXiv is drawing a line between author-responsibility and AI-generation that goes beyond the old 'we don't police tool use' posture.

00:05:05 The policy matters because the volume of AI-assisted submissions has been increasing, and the community has been asking for clearer standards. The hallucinated references point is the right one — it's not about using AI to draft. It's about whether authors verified what the AI produced.

00:05:26 I haven't seen another preprint server announce a comparable policy. This is a concrete step, and the one-year ban is a real deterrent.

00:05:36 A different tension today — between safety narratives and compute economics at Anthropic. Curtis Pyke published a long investigation into why Anthropic's Claude Mythos Preview is behind a closed door. Mythos is the company's most powerful model, available only through a 40-organization security research program called Project Glasswing.

00:06:00 Pricing during the preview is $25 per million input tokens and $125 per million output — substantially higher than Claude's standard tiers. The safety case, per Anthropic's Frontier Red Team: Mythos crossed a threshold where it can discover and exploit zero-day vulnerabilities autonomously.

00:06:21 The restricted release is designed to buy time for defenders. That's the narrative. But the compute story is concrete. Within days of the Glasswing launch, Anthropic announced compute deals with Google-Broadcom for 3.5 gigawatts of capacity starting in 2027. They also signed a separate deal with CoreWeave for capacity later this year.

00:06:45 Reuters reported Anthropic was exploring designing its own AI chips — a half-billion-dollar effort explicitly tied to a broader shortage. Pyke weighs both explanations. Neither fully explains the other. The core tension here is that safety constraints and compute economics aren't competing stories.

00:07:06 They're simultaneous constraints, and they shape each other. If Mythos needs that much compute to run, even defensive use becomes expensive. The $100 million in credits Anthropic committed to Glasswing participants is a signal of strategic investment, but it's also a signal that running these models costs real money.

00:07:29 The question is whether the safety argument and the compute argument converge — whether the models that are hardest to run are the ones that are hardest to keep safe. That's worth watching as the frontier moves.

00:07:44 Finally, a couple smaller stories. Osaurus, an open-source Mac app that combines local and cloud AI models, hit its first anniversary with over 112,000 downloads. Co-founder Terence Pae started building it after customers asked why they should pay for tokens if they could run AI locally.

00:08:06 The app requires 64 GB minimum RAM, 128 GB recommended for larger models like DeepSeek v4, and supports everything from MiniMax M2.5 to GPT-OSS to Anthropic's models. The interesting part about Osaurus is the hybrid approach — you run your own local models, keep your files on your hardware, and fall back to cloud providers when needed.

00:08:31 That pattern of local-first with cloud extension is becoming a real architecture for people who care about privacy and cost control. On the hardware side, NVIDIA reportedly prepares an RTX 5090 price increase due to rising GDDR7 costs. The RTX 5000 Pro 48GB arrived and early reviews are positive.

00:08:54 Local inference is getting cheaper per watt, but the hardware floor is still moving up. And over in robotics, Figure AI's humanoid kept working for over 30 hours without breaks. It's a long run, and it highlights a different kind of capability — endurance, not intelligence.

00:09:14 The gap between robot uptime and human needs is still wide, but narrowing. That's the local pass for Friday. — Seln.