A new EMNLP benchmark puts a hard number on the cheapest optimization: shrink the output, not the prompt. Meanwhile, an agentic adoption index reveals delegation is concentrated in mid-wage jobs, not the usual suspects.
Swift-Image packs text-to-image generation plus single- and multi-image editing into a single 6B set of weights, scoring top among open-source models on editing benchmarks with just 243K GPU-hours of training. Structural pruning yields a near-lossless 3B variant that still leads. Speculation is that Alibaba may ship it as a fully open model.
A new EMNLP benchmark (CAVEWOMAN) put hard numbers behind a hunch: telling an LLM to be concise cuts API cost about 1.5x on average, up to 3x, with accuracy nearly flat. Compressing the input prompt does the opposite, since the model rambles to fill gaps, costing up to 96% more while answering worse. Across 9 models, 8 datasets, and 11 languages, the verdict is blunt: shrink the output, not the prompt.
Existing AI-exposure models measure where AI could work, not whether workers have adopted it. This paper introduces "delegated exposure," an Agentic Adoption Index built by matching 53,000 shared agent skills from Manus against 18,000 O*NET task statements. The result: real delegation concentrates in different occupations than pre-AI risk frameworks predicted, and adoption peaks in mid-wage bachelor-level work, declining among the most educated.
Safety alignment in LLMs is skin-deep: refusal triggers live at the final layers while pretrained harm knowledge stays buried. This paper shows how "semantic camouflage", harmful intent wrapped in benign creative-writing prompts, becomes mathematically indistinguishable from safe queries late in the network, yet retains a detectable early-layer harm signature. Latent Intent Verification, a lightweight probing defense, exploits that, beating standard guardrails by 20-50% across Phi-3, Qwen2.5, and Gemma-2b with no retraining.
Xiaomi's AI Cube is a 150W mini-PC prototype that crams three of its own in-house chips (O3 mobile SoC, O100 AI accelerator, D100 automotive processor) into one desktop. The O100 alone hits 1.22TB/s of near-memory bandwidth, enough to run 120B-parameter models locally while switching between fast and slow systems. No price or release date yet.
Google now includes Antigravity in eligible Gemini Enterprise (Standard and Plus) subscriptions, letting orgs hand every developer autonomous coding agents under one billing and governance umbrella. New IDE extensions land for VS Code, Visual Studio, JetBrains, and Zed, with Workforce Identity Federation login so engineers skip API-key setup. Admins get sandbox policies, pooled token budgets, overage caps, and one-toggle audit logging. AirAsia already generates over half its production QA code this way.
Reverse-engineered Claude Desktop builds reveal an unreleased macOS feature, codenamed Parka, that records meetings, transcribes them speaker-by-speaker, and turns action items into tasks Claude agents can actually execute. Meeting summaries are commodity; Parka's edge is handing the follow-up work straight to Claude Code and Cowork. It ships disabled behind a bare loader stub, so for now this is a leak of what's coming, not a live feature.
The UK's competition regulator is weighing rules that could put ChatGPT and Perplexity on Google's default-search-engine choice screen. Publishers are pushing back: they've watched traffic crater since AI Overviews, and fear that making an AI assistant a one-tap default would accelerate the collapse. Trade bodies warn the wording of the CMA's conduct requirements risks enshrining AI as search. Google's own screens currently don't offer AI options at all.
Turkey's censorship watchdog has pulled at least 12 posts from inside Grok, xAI's chatbot, on national security grounds. The orders rest on Article 8/A of Internet Law 5651, the same provision Ankara uses to silence journalists and X accounts. It could be the first real test of how states police AI-generated speech, and of whether xAI fights or quietly complies.
The machine holding a model's weights sits right next to the vulnerable code that serves it. Inference servers like vLLM and SGLang are big, buggy C/Python surfaces, and vLLM even shipped eval() on tool-call parameters. Since agentic harnesses execute model output on separate machines, a compromised inference host becomes the real crown jewel, not the model itself. The essay maps the attack chain and what defenses would blunt it.
You’re caught up.
That was all 10 stories in the 24 August issue.