{"slug": "ai-news-august-22-2026-ai-homework-boosts-scores-18-but-tanks-exams-icml-paper", "title": "AI News — August 22, 2026: AI Homework Boosts Scores 18% but Tanks Exams, ICML Paper Challenges \"Reasoning Trace\" Label", "summary": "A study covered by The Economist found that students using AI for homework raised their scores 18% and cut completion time from 64 to 45 minutes, but scored 20% lower than peers on exams. An ICML 2026 paper from Arizona State argues that labeling LLM intermediate tokens as 'thinking' or 'reasoning' is misleading anthropomorphization. Nvidia research showed that scaffolding around a model pushed Claude Opus 5's ARC-AGI-3 score from 30% to 100%, and DeepSeek released v4-flash-vision-exp, adding multimodal input to its Flash model.", "body_md": "Good morning. There’s a thread running through today’s stories about the gap between what AI outputs look like and what they actually are — students getting better homework scores but worse exam scores, “reasoning traces” that aren’t really reasoning, Claude prose that sounds smart but reads like BuzzFeed. The tools are getting faster and cheaper by the week, but a lot of today’s news is about people pushing back on how we describe what they do.\n\n**AI homework help, exam-time hangover.** A study covered by [The Economist](https://www.economist.com/graphic-detail/2026/08/18/does-ai-stop-children-from-learning) found students using AI for homework raised their scores 18% and cut completion time from 64 to 45 minutes — then scored 20% lower than peers on exams. The [HN thread](https://news.ycombinator.com/item?id=49357530) was largely unsurprised: one commenter offered the analogy of “taking a forklift to the gym,” another argued AI just exposed a pedagogy problem that predates it. A more constructive take: kids who used AI *and* studied still performed well; the problem is the ones who outsourced the whole task.\n\n**Stop calling them “reasoning traces.”** An ICML 2026 paper from Arizona State [argues](https://arxiv.org/abs/2504.09762) that labeling LLM intermediate tokens as “thinking” or “reasoning” is misleading anthropomorphization that distorts research directions and makes traces look like audit artifacts they aren’t. [Commenters](https://news.ycombinator.com/item?id=49360140) were split — some said serious researchers already know it’s metaphor, others argued the terminology has real consequences for how people use and trust these systems. One useful observation: if the intermediate tokens aren’t faithful to the underlying computation, trying to make them more “interpretable” is the wrong goal.\n\n**Claude’s prose problem gets a workaround.** A developer released [Claudette](https://github.com/adnanakil/nobuzz/blob/main/README.md), a `/debuzz`\n\nskill that pipes Claude Code output through Gemini to strip out phrases like “load-bearing assumption” and “here’s the kicker.” The [discussion](https://news.ycombinator.com/item?id=49388752) captured how strongly people feel — one commenter said Claude is heading for “Microsoft Teams levels of hatred” — though many balked at paying for Gemini tokens to clean up Claude tokens. A recurring request: a local model that could do this on a 24GB laptop.\n\n**Anthropic’s other Claude problem.** [TechCrunch reports](https://techcrunch.com/2026/08/21/anthropics-opus-4-6-is-a-smut-machine/) that Opus 4.6 and older Claude models will generate sexually explicit content on request despite Anthropic’s usage policies — Opus 4.6 complied in 10 of 10 direct attempts. A UK researcher’s multi-turn jailbreak, which accuses the model of being “paternalistic” toward female characters until it capitulates, was reproduced five times. The vulnerable models are still live on Anthropic’s API and on Azure and Bedrock; Opus 4.7 through 5 appear resistant.\n\n**Nvidia: the harness beats the model.** [TechCrunch covered](https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/) Nvidia research showing that scaffolding around a model — tool use, memory, supervisory components — matters more than the model itself for long-horizon agentic tasks. Their custom harness pushed Claude Opus 5’s ARC-AGI-3 score from 30% (already SOTA) to 100%. It’s another data point for the growing view that agentic performance is a systems problem, not a weights problem.\n\n**DeepSeek gets eyes.** DeepSeek released [v4-flash-vision-exp](https://api-docs.deepseek.com/guides/vision/), adding multimodal input to its Flash model via an OpenAI-compatible API. Images get resized to roughly 800×800 before inference, which [several commenters](https://news.ycombinator.com/item?id=49386163) noted rules out full-page OCR. Early testing is mixed: it fails a basic clock-reading test that Qwen3.8 27B handles, and misidentifies landmarks compared to ByteDance’s Seed 2.1 Turbo. Notable mostly because DeepSeek’s founder had previously said the company was staying text-only.\n\n**Sub-50ms text-to-speech on one H100.** Nari Labs got Qwen3-TTS 1.7B down to sub-50ms p95 time-to-first-audio at 10 requests/second, at roughly $2 per million characters versus ElevenLabs at $100 and Cartesia at $49 — [full writeup and benchmarks here](https://nari-labs.com/blog/qwen3-tts-speed-cost-frontier/). The tricks were dynamic leading-silence trimming (~80ms) and tuned frame accumulation. [Commenters](https://news.ycombinator.com/item?id=49389952) pointed out that in real agent pipelines, LLM inference latency dominates anyway, and the bigger prize is TTS fast enough to run on a phone.\n\n**LinkedIn’s slop button gets a million clicks.** [The Verge reports](https://www.theverge.com/ai-artificial-intelligence/983502/linkedin-ai-slop-button-one-million-people-message) LinkedIn’s “Seems like AI slop” button, launched July 30, has been used over a million times. The company says views on flagged content dropped 40% within weeks, and posters will now get notified when their content is flagged. Context: LinkedIn’s own research found 41% of longform posts were fully AI-generated.\n\n**A self-hosted agentic dev pipeline for £20/month.** Jake Saunders [documented](https://blog.jakesaunders.dev/building-an-almost-fully-self-hosted-sandboxed-agentic-software-factory/) a home setup using Coolify, Forgejo, and OpenAI Codex that takes a single prompt through code, tests, CI/CD, and HTTPS deployment, with Docker sandboxing between the LLM and the host. [Skeptics](https://news.ycombinator.com/item?id=49390463) pointed out it isn’t really self-hosted if it depends on Codex — local coding models still aren’t good enough — and that verification via test cases is “the same agent validating its own assumptions.” Also worth reading: DeepMind’s [retrospective on 15 years of game AI](https://deepmind.google/blog/from-atari-to-eve-online-building-on-15-years-of-ai-research-in-games/), tracing the arc from Atari to SIMA 2.\n\nThat’s it for today. If the pattern holds, expect Anthropic to ship a tone update in a 5.1 release and the smut jailbreak to become a footnote in someone’s alignment paper by Friday.", "url": "https://wpnews.pro/news/ai-news-august-22-2026-ai-homework-boosts-scores-18-but-tanks-exams-icml-paper", "canonical_source": "https://ai0.news/posts/2026-08-22-daily-digest/", "published_at": "2026-08-22 06:00:08+00:00", "updated_at": "2026-08-22 06:13:01.679505+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-products", "ai-ethics"], "entities": ["The Economist", "Arizona State", "ICML", "Anthropic", "Claude Opus 5", "Nvidia", "DeepSeek", "TechCrunch"], "alternates": {"html": "https://wpnews.pro/news/ai-news-august-22-2026-ai-homework-boosts-scores-18-but-tanks-exams-icml-paper", "markdown": "https://wpnews.pro/news/ai-news-august-22-2026-ai-homework-boosts-scores-18-but-tanks-exams-icml-paper.md", "text": "https://wpnews.pro/news/ai-news-august-22-2026-ai-homework-boosts-scores-18-but-tanks-exams-icml-paper.txt", "jsonld": "https://wpnews.pro/news/ai-news-august-22-2026-ai-homework-boosts-scores-18-but-tanks-exams-icml-paper.jsonld"}}