Julien Simon

I read filings, price sheets and model cards, then set one dated fact against another. No hype, primary source attached. Three recent findings from The AI Realist 🧵

Open-source modules now unlock 2x the training math on GeForce cards. That's real. @nvidia's own whitepapers show the speed was in the silicon all along, but a setting held it back. Here is the paradox: the math really doubles, yet your training job doesn't run twice as fast, and token generation doesn't move at all. I rented the cards, measured both sides of NVIDIA's line, and priced what each one actually delivers. Where the 2x goes, and why the capped card was already the better buy, is in the piece. airealist.ai/p/nvidia-geforc…

I ran Parallel's Decisions API on true/false claims. It returns a probability and citations instead of a bare yes or no. Anthropic IPO, one call: ✅ "Filed an S-1": 0.97 ✅ "Shares trading today": 0.12 ✅ "Going public in 2026": 0.98 Hugging Face, one call: ✅ "Headquartered in Paris": 0.12 (it's New York) Cost by Parallel's meter: $0.001 per call. Latency: 0.66 s, including network. Citations include the SEC filing. The 0.12 on trading is not near zero, so low-confidence answers still need a human. #ParallelPartner

Pretty cool. You can follow a vulnerability across linked repos, then replay the exploit in a sandbox to prove it's real.

Every Mac with @Apple Intelligence now carries a 20-billion-parameter language model, and you can call it from a terminal. So what is it, exactly? How is it built, who built it, and what is it good for? Apple publishes no model card, so I built one from the outside: file reads, memory, power, token counts. What I found is beautiful engineering. The model lives in flash storage. For each prompt, it wakes a few experts and runs them on the Neural Engine. After a long request, it held 2 GiB of memory. An open model of the same class, built the usual way, held 15. @Google's part is in the training. The engineering is Apple's. Every measurement, script, and dataset is on GitHub, so you can rerun it on your own Mac. ➡️ Substack: airealist.ai/p/exploring-app… ➡️ Github: github.com/juliensimon/apple…

Moody's maintient la France à Aa3. Les investisseurs la classent en BBB-. Quand l'agence et le marché divergent de 5 crans, c'est le marché seul qui paie son erreur.

Moody's holds France at Aa3. Investors hold France at BBB-. When the agency and the market disagree by 5 notches, the market is the only one paying for being wrong.

Does paying for @OpenAI embeddings beat a free 0.6B model on your laptop? I tested it on Stripe's docs in English, French, German, and Spanish. 1,626 questions, five models, a 95% interval on every gap. English, right passage first: OpenAI 64.9%, Octen 65.9%, plain e5-large 66.7%. The test cannot tell them apart. In German and Spanish, the old free e5-large is ahead of OpenAI. youtu.be/mYgYUDHaJWA github.com/juliensimon/strip…

Appel à @AmericanExpress il y a 2 semaines: Bonjour, je vous ai envoyé mes nouvelles coordonnées bancaires il y a une semaine. Vous les avez bien notées ? "⁠⁠Ah non, il manque le prénom." Pardon? "Oui, votre nouvel IBAN dit J SIMON" "⁠⁠Mais c’est bien mon IBAN, mon nom et mon adresse." "Oui, mais il manque le prénom. On ne peut pas le prendre." ⁠⁠Et donc ? "Vous devez demander à votre banque de mettre à jour l’IBAN." C'est une plaisanterie ? Il n'y a aucun autre moyen ? "Non." ⁠Je suis client depuis 25 ans. Un petit effort, s'il vous plaît ? "⁠⁠Désolé, c'est impossible." Parfait, je ferme mon compte à la fin du mois. "OK." Je les déglingue sur l'enquête de satisfaction reçue par mail, comme il se doit. Une semaine plus tard, sans aucune information de leur part, le prélèvement a lieu sur le nouveau compte, heureusement approvisionné. 👏👏👏 Et donc, je ferme mon compte à la fin du mois.