The EU's new rules require telling users when they are interacting with AI and when they are viewing content that AI generated or edited — and this has given rise to concerns about "disclosure fatigue." While platforms are strengthening their response to low-quality AI content, engineers in the field are deep in debate, re-examining the cost and speed of inference, the design of agents, and the very foundations of reinforcement learning. Here are today's five topics.
Snap Bans AI-Generated Videos from Spotlight; LinkedIn Adds an "AI Slop" Report Button
Snap has indicated a policy of shutting AI-generated videos out of Spotlight, its short-video feed, saying it wants to keep the feed focused on human creativity; content edited with Snapchat's own first-party AI tools will be allowed as an exception. Meanwhile, LinkedIn has rolled out a dedicated button for reporting low-quality AI-generated material.
The spread of low-quality AI-generated content known as "AI slop" has become a headache for platform companies. Both Snap and LinkedIn have moved to respond by narrowing the intake points on the platform side. The difficulty of drawing lines — what to count as "AI-generated" and how to treat edits made with first-party tools — has also come into relief, suggesting that both policy and technology will require ongoing adjustment.
The Trap of Inference Speedup "Combinations" That Backfire
"If you enable both speculative decoding and prefix caching, speed should improve multiplicatively" — this naive expectation is often betrayed in practice.
A real bug report filed on vLLM's GitHub describes how using prefix caching with MTP (Multi-Token Prediction)-style speculative decoding enabled halved cache hits from 32,768 tokens to 16,384 tokens and significantly worsened TTFT (time to first token). It's a case of hard-won optimizations interfering with each other and dragging each other down.
Both are staple acceleration techniques for LLM inference, but "stacking things that sound good" does not necessarily make things faster. You need to understand how cache boundaries mesh with the behavior of speculative generation, and design the combination on that basis.
Is "Strong Model for Design, Cheap Model for Implementation" Still the Right Answer?
The phase-based division of labor — "let a high-performance strong model handle design and a low-cost cheap model handle implementation" — is still effective given the right conditions. However, an analysis argues that it is hard to call this a best practice that can be recommended to everyone.
What deserves attention is that the official documentation for Claude Code and Codex decides what to delegate based not on phase names but on "self-containedness of the work," "continuity of context," "write contention," and "verifiability." Work that can compress a large amount of information into a short summary — investigation, log analysis, test execution, review — suits subagents. Design and core implementation, on the other hand, involve a steady stream of specification decisions and design changes during the work, so keeping the same main agent on them tends to run more smoothly.
In short, the essence is not "which phase it is" but "whether the work can be cut apart and delegated." If you leave implementation to a separate agent, it is apparently more realistic to carve out the work in delegable units rather than throwing the entire phase over the wall.
Pasting Long Text into an AI Costs 4.2x More, but Tokens Only Grow 1.5x — and Wait Time Doesn't Scale Proportionally
"Paste a whole long document into an AI and it gets that much slower and more expensive" — an article that put this assumption to the test with real measurements in a cloud sandbox makes for an interesting read.
The results split cleanly between what grew and what did not. The bill rose in proportion to character count, at about $0.0614 per additional 10,000 characters. Yet the token count increased only 1.5x even when the pasted text was multiplied 346x. Moreover, wait time did not scale with input length — the conditions overlapped within the same band, and the median even dropped partway through.
The naive proportional relationship "characters = tokens = cost = wait time" does not hold. The intuition (long text = expensive and slow) is mostly correct on the cost side, but the behavior of token counts and latency is not so easily tamed.
Long-CoT Reinforcement Learning Re-Asks "Which Tokens Deserve Credit"
A research group from Nanjing University and Shanghai Jiao Tong University has confronted head-on a fundamental problem in reinforcement learning that assumes long chains of thought (CoT): "When the model arrives at the correct answer, which tokens should be credited for the contribution?"
The widely used GRPO distributes the response-level reward equally across tokens, which is not necessarily correct. Self-distillation-style methods take "the likelihood shift when privileged information is provided" as a cue, but according to the authors' counterfactual analysis, that shift often points in a direction unrelated to the correctness of the answer. Large shifts concentrated on interchangeable, surface-level tokens such as "But," "So," and "Let," while problem-specific reasoning tokens were rather insensitive.
In response, a method called "CSCR" has been proposed that redistributes rewards at the token level. It is an attempt to improve the training efficiency of reasoning models from the ground up, and it once again throws into sharp relief the difficulty of "credit assignment" in reinforcement learning.