Frontier Watch

I put eighteen empty web addresses online. Nine got their own security certificate, which writes the name into the public Certificate Transparency logs, and all nine were visited within a day, the first within one to four seconds of the log entry. Eight of the nine kept out of the logs by a wildcard certificate got nothing for a day. What the scanners went looking for, how fast the AI crawlers came, and what to do if you launch websites. Site: https://ctlogs.novcog.us.com/ · Paper: https://doi.org/10.5281/zenodo.24203364 · Video: https://youtu.be/NmFBuLsshBg

A new paper claimed reward training doesn’t teach AI anything new. I re-ran its test, wrote it up as a pre-print, and found the improvement it set out to study had already happened before its first measurement: the model was trained in two rounds, and the paper only watched the second. I also ran the test it proposed and didn’t run, whether a mistake-checker built before training keeps working after it. It does. This corrects episode 48. The pre-print: https://doi.org/10.5281/zenodo.24096296 · The data and code: https://doi.org/10.5281/zenodo.24096294 · The video: https://youtu.be/ays-X4dU6PU · Full write-up: https://rlvr.novcog.us.com/dataset

Correction, 10 November 2026: this episode says the gain lands in the first 200 steps of reward training. It doesn’t. The paper’s snapshots come from a second round of training that starts from an already trained model, and the gain was made in the first round. The corrected story is episode 49, and the pre-print with all the data is at https://doi.org/10.5281/zenodo.24096296 . A new workshop paper, RLVR Sharpens the Pencil, Not the Mind (arXiv:2611.02574), argues that reward training only makes a model lead with an answer it already had. I checked it against its own tables and code, re-ran its accuracy test from the step it skips, and read the papers around it for what to do with the result and what to test next. Full analysis, sources and the re-run: https://rlvr.novcog.us.com/ · Video: https://youtu.be/ygoyREfcNVQ

A viral post quotes Andrej Karpathy saying today's models waste capacity remembering things and that a one-billion-parameter model can beat most of them. Most of that is his, from the No Priors podcast in September 2024; the 90% and the comparison with today's models are not. webAI's TwIL-8B is held up as proof: its 93/92/90 percent-of-frontier scores are webAI's own, its licence is non-commercial research and education only, and in our spot check on a Mac Studio it passed 164 of 168 checks but took twice as long per answer as DeepSeek V4 Flash. Full write-up: https://cogcore.novcog.us.com/ · Video: https://youtu.be/8wW9LZ1UfuE

I ran WiCi's open-source InferSSD engine and the standard llama.cpp engine on the same 500 GB DeepSeek V4.1 Flash model, on our Mac Studio with 256 GB of memory, using WiCi's own test script. llama.cpp wrote answers three times faster (10.95 vs 3.45 tokens a second); InferSSD read the prompt as fast or faster; its 200 GB setting never ran. WiCi's 48 GB laptop result is a different machine and is untouched. Every number and the red team: https://inferssd.novcog.us.com/ . Video: https://youtu.be/c5GGiiNGL6A . Want this run on your own stack? https://novcog.com/contact

NetworkChuck ran a 748-billion-parameter DeepSeek model on WiCi One, a box with a 32 GB RTX 5090 that streams most of the model off an SSD, at about 8 pieces of text a second. His video is sponsored by WiCi, and he says so. I checked 21 of its claims against DeepSeek's model card, WiCi's open-source InferSSD repo and Apple's spec pages: 12 hold, 6 hold with a caveat, none fail. Then I ran what I could on our own Mac Studio (M3 Ultra, 256 GB): the same coffee questions on the smaller DeepSeek V4 Flash at 22.5 a second, our model drive at 2.6 GB/s for big reads and 34 MB/s for tiny ones, a model switch in 0.69 seconds, and an answer that used 18 kilobits a second on the network. Written analysis: https://familygpu.novcog.us.com/ · Video: https://youtu.be/fXe88jyy7lk

Three AI models charge the same price per token, and one finished task costs $1.27, $3.68 or $10.02. I checked what AI agents really cost per task on every vendor price page and on Artificial Analysis's new Intelligence Index v5.0 (29 Oct 2026): GPT-6.1 Terra and Sol, Claude Opus, Sonnet and Haiku, Gemini 4 Argon (doubling 1 December), Kimi K3.1 Pro, DeepSeek, MiniMax and GLM. Every price and score, with sources: https://agentcost.novcog.us.com/ . Video: https://youtu.be/3hdXa54Vb7w

A semantic cache answers a repeated question without calling the model. It can cut an LLM bill on repetitive traffic, and it can hand a customer the wrong answer with total confidence. We checked Redis LangCache against Redis's own documents (still in preview; "Save 90% on API costs"; the Mangoes.ai 70% quote; Redis's own savings formula), read the research, and measured the threshold on 243 support-question pairs: cancel vs pause scored higher than a true rewording. Sources and our test data: https://semanticcache.novcog.us.com/ . Free traffic review: https://novcog.com/contact/ . Video: https://youtu.be/XW-VsLfYiqg

Moonshot Kimi K3.1 Pro ranks 4th of 215 on Artificial Analysis and topped an independent Terminal-Bench 4.0 run, but at launch it is API-only with weights promised by Nov 4. Checked against the live API, independent evals and community reaction: speed, cost per task, the cyber-data removal, and what it means for open and closed models. Sources: https://k31pro.novcog.us.com/ Video: https://youtu.be/tNEDJx2QEV8

A veteran SEO says generative engine optimization has been mathematically disproven. I read every study he cited. The critique of the 2023 GEO paper mostly holds, but a sentence quoted from one study is not in that study, a lab experiment is described as the live web, and the Toronto study it leans on found 97% of the pages ChatGPT retrieved were not in Google top 50. Every claim checked at https://geocoffin.novcog.us.com/ and the video is at https://youtu.be/6xIpBRgqaWc