Weekly YouTube Digest - Jul 13-19, 2026

This week's fast take

You aren't using Codex like me...

• His routing rule is simple: send the hardest work to GPT-5.6 Soul, use Luna for simpler tasks, and treat thinking effort as a budget to tune; he says this is his own operating preference, not a neutral benchmark result 1.
• He uses agents.md to encode project rules, including which model, thinking effort, thread, and folder should be used for a command such as deploy; he also recommends revisiting stale rules after a model release 1.
• Threads become a lightweight delegation system: one thread can start work in another, while a main thread supervises several parallel tasks; the video demonstrates the pattern rather than measuring its reliability 1.
• The broader workflow includes browser actions, reusable skills, and long-running goals or loops that keep iterating on a benchmark or codebase for hours; the benchmark improvements described are the speaker's own workflow anecdotes 1.
• The most important operational advice is defensive: use approval prompts and pre-tool hooks to block destructive commands before execution, especially if an agent has broad filesystem or internet access 1.

Did Kimi K3 really beat Fable?

• Berman cites an Arena AI front-end benchmark where Kimi K3 is reported at 76% versus Fable 5 at 63%, presenting it as evidence that Kimi can lead on a coding-oriented task 2.
• He also discusses Deep Suite cost-per-task and completion results, saying Kimi K3 Max lands near GPT-5.6 Soul at roughly $4.70 while using more tokens for the same task; those figures are video claims, not independently re-run here 2.
• A Next.js evaluation and an internal writing benchmark are used to support the same story: Kimi performs strongly on agentic software work and can be cheaper than the model it replaced 2.
• The practical drawback is speed and token consumption, so a high score does not automatically translate into lower latency or lower total cost in a production workflow 2.
• Berman's conclusion is measured beneath the headline: Kimi K3 looks like a formidable open-weight competitor, but benchmark saturation, real-world validation, and an unresolved distillation allegation all limit how far the result should be generalized 2.

Claude Just Revealed AI's Biggest Problem

• The study split 52 junior software engineers into an AI-assisted group and a no-AI group; the AI group finished about two minutes, or roughly 8%, faster, but that speed difference was not statistically significant 3.
• On a follow-up quiz, the AI group scored 50% versus 67% for the no-AI group, a statistically significant gap in the study summarized by the video 3.
• Debugging was the clearest weakness: generating code with assistance is easier than diagnosing a failure when the underlying logic was never learned well enough to inspect 3.
• The proposed response is practical: use AI mainly for work you already understand, ask it to teach when you are learning, and try to debug a broken result yourself before asking the model for a fix 3.
• The video stresses the limits: one small group of mostly junior developers, one Python library, one short task, one quiz, and a chat assistant rather than a full agentic coding system do not settle the broader question 3.

Anthropic Found Something That Shouldn't Exist

• The task sounds trivial: decide whether a word such as "aluminum" fits within a page, but the model must infer character length and page position from tokenized input rather than receiving either value explicitly 4.
• The video describes sparse, neuron-like features that activate for positions along the line or near the page boundary, comparing them to place and boundary cells in biological systems 4.
• The internal counter is not presented as a simple integer register: it tracks token position and uses a curved, rippling representation that separates nearby values into more distinguishable channels 4.
• The important interpretability move is causal reconstruction: researchers can identify the internal features and connect them to a behavior the model learned during training, rather than stopping at a correlation between an activation and an answer 4.
• The result is evidence that models can develop compact, task-specific algorithms inside their representations; it is not evidence that one toy task establishes general intelligence or that every model will form the same circuit 4.

Learning Genetic Perturbation Effects at Single-Cell Resolution for Virtual Cells

• The motivating data come from destructive experiments such as Perturb-seq and optical pool screens, where control and perturbed cells are not paired and exhaustive measurement is expensive 5.
• The presented MORPH-style approach maps control cells into a latent space, combines them with perturbation representations, and uses a distribution-level loss such as maximum mean discrepancy instead of relying on cell-by-cell reconstruction error 5.
• Prior biological knowledge is central: perturbation embeddings can come from measured co-expression, single-cell foundation models, gene descriptions, or dependency maps, allowing the model to reason about unseen genes instead of treating every perturbation as an unrelated one-hot label 5.
• The talk reports gains over mean baselines, GEARS, and other embedding methods on genome-wide prediction, as well as improved prediction of infection-stage proportions in an Ebola optical-pool setting 5.
• The most tangible application is experiment design: in a cancer-immunotherapy collaboration, the model nominated two T-cell regulator targets beyond 73 tested genes, but the generalization depends on causal assumptions, informative priors, and enough relevant perturbational data 5.