"Iterative Finetuning is Mostly Idempotent" If a model has a behavioral tendency and you train the next generation on its outputs, does the tendency snowball? This paper actually finds that it's mostly no. So they seed models with traits like sycophancy and misalignment, then
LLMS
-
Demystifying Manifold Constraints in LLM Pre-training
By
–
"Demystifying Manifold Constraints in LLM Pre-training" LLM pre-training depends on a stack of heuristic tricks, eg. normalization layers, weight decay, and optimizer tuning. This paper suggests that by constraining weight matrices to geometric manifolds like spectral or
-
Excitement for upcoming AI models: GPT 5.6, Opus 4.8, Kimi, GLM
By
–
Don’t know if I’m more excited for gpt 5.6, opus 4.8 a new Kimi or GLM model. What a freaking amazing time to be alive
-
The Economic Barrier Between AI Model Ownership and Wrappers
By
–
Unless you own the models it is a wrapper, and training models costs $100M-$1B
-
The High Cost of Training AI Models
By
–
Yep unless you own the models it is, and training models costs $100M-$1B
-
Excitement for GPT-5.5 and next model hype
By
–
GPT-5.5 is the goat right now. Freaking excited for the next model. And Sama is great hyping it up 😉
-
OpenAI Plugin Demonstrates Multi-Agent AI Workflow with Claude Code and Codex
By
–
OpenAI launched a plugin so that Claude Code can use Codex And it just demonstrated why the future is not a single AI agent. It's several working together. It's not smoke. It's how LLMs work. The flow is simple: • Claude writes the code.
• Codex reviews it in an -
Québec AI Sovereign Manifesto: Building AI‑First Enterprises
By
–
The http://
QUEBEC.AI Sovereign Manifesto. The next great enterprises won’t just use AI. They become AI‑First intelligence organizations: – Agents discover.
– Validators prove.
– Memory preserves.
– Governance controls.
– The enterprise compounds. Québec’s sovereign AI