Multimodal

Google added an AI generation button to Google Earth on July 30, 2026, letting users create synthetic images on real satellite basemaps with Nano Banana 2, then pulled it within a day. The core issue: screenshots shared on social media lost AI watermarks and metadata, but Google Earth's 20-year reputation as a 'window on reality' traveled with them. OSINT analyst Henk van Ess generated fake craters, flooded landmarks, and destroyed sites, noting the fakes inherited the map's credibility. The article contrasts Wikipedia banning AI edits, Snapchat embracing AI lenses, and LinkedIn removing AI writing while adding a slop-report button, arguing that platform attitudes hinge on the interface promise made to users—factual archive vs. playground. Provenance tech like SynthID and C2PA degrades across platforms; labels alone barely shift user belief; outright bans and combo strategies (labeling + demonetization + downranking) are the main responses.

Alexi Gladstone introduces Explorative Modeling (XM): generate K candidates per step, train only on the best. This adds a third pretraining axis beyond data and parameters. More exploration monotonically improves image, video, and language models, with gains growing at scale—7%→36% with more data, 13%→23% with more parameters. XM achieves 6.2× sample efficiency, 4.1× FLOP efficiency, and 47% better parameter efficiency. As an end-to-end generator, XM matches diffusion on control tasks using up to 256× less inference compute. The post does not disclose specific model names or training costs.

Google pulled the 'AI satellite view' feature from Google Earth on July 31, days after launch. The tool was meant to sharpen blurry satellite images with generative AI, but users found it was inventing buildings, roads, and vehicles on real terrain. Google confirmed the feature 'did not meet quality standards' and has taken it offline. No timeline for a fix was given, and the post doesn't spell out the exact conditions that triggered the hallucinations. This is another case of generative AI failing in a non-fiction product—especially damaging for a map tool that relies on factual accuracy.

Google added Nano Banana 2 to Google Earth on Thursday, letting users prompt-generate images over real satellite maps. BBC journalists immediately flagged it as a misinformation risk. Google pulled the feature Friday, saying some generated screenshots violated its policies and stronger guardrails are needed before a re-release. The post doesn't detail what those guardrails are or when the feature might return.

Google Earth's new Dream House feature generates aerial-style images from text prompts. Security researcher Henk van Ess showed that adding "satellite view" to a prompt bypasses filters and produces convincing fake satellite imagery. Google applies watermarks and content restrictions, but van Ess still got results using variants like "aerial view." Reverse image searches then indexed those fakes as real locations in third-party map databases. The core risk isn't the generated image itself—it's that these images can slip into workflows that rely on overhead views for news verification, insurance claims, or intelligence analysis.

Minnesota's law banning nudification apps is about to take effect, and xAI filed a last-minute lawsuit to block it. xAI argues the law is overbroad and would restrict Grok's image generation, violating First Amendment free speech. In the filing, xAI describes Grok as an opinionated, sarcastic AI assistant whose explicit images are a form of expression. The state attorney general counters that the law only targets non-consensual fake nudes and has nothing to do with free speech. The case has just been filed and hasn't been heard yet.

Illustrators, authors, and musicians are filing copyright lawsuits against Google, Meta, Anthropic, and others. The piece tracks recent case updates: some courts have denied the tech companies' motions to dismiss, letting the suits proceed. Artists feel more optimistic about their legal odds than before, but remain pessimistic about AI's overall direction. The post does not disclose specific damages or settlement details.

Sebastian Raschka breaks down Kimi K3, the largest open-weight model at 2.8T params, scaled from last year's 48B Kimi Linear. The design prioritizes inference efficiency: LatentMoE compresses large linear layers, multi-head latent attention and Delta Attention replace standard attention, and RoPE is dropped entirely for NoPE. The only non-efficiency tweak is attention residuals, which add 4% training cost for consistent small gains in validation loss and downstream performance. Native multimodal support is also included.

A Verge investigation found that Hugging Face hosts numerous models capable of generating nude images, many targeting women and children. These models are disguised as 'clothing change' or 'fashion editing' tools, requiring only a single photo to produce a nude output. The platform currently implements almost no safeguards at a system level and does not proactively scan uploaded models. Hugging Face says it relies on manual review of reports, but the post does not disclose the size of the review team or response times. I'd take 'zero safeguards' with a grain of salt—the platform does have a content policy, but enforcement appears far behind the pace of abuse.

Moonshot AI open-sourced Kimi K3 weights, a tech report, and three infra projects in one drop. K3 is a 2.8T-parameter MoE model with native vision and a 1M-token context window. The team claims 2.5× scaling efficiency over K2.5. The three infra releases—MoonEP, FlashKDA, and AgentEnv—aren't detailed in the snippet, but the names point to expert parallelism, attention acceleration, and an agent environment.

Kimi open-sourced K3, its strongest model: a 2.8T-param MoE with native vision and a 1M-token context window. The new architecture claims 2.5× intelligence per unit of compute. Weights, high-performance attention kernels, an MoE communication library, and a large-scale agent runtime are all released. The post doesn't disclose training data, benchmark scores, or the license.

Bonsai 27B compresses Qwen3.6-27B to ~1.125 bit/weight, fits a ~3.9 GB working set on iPhone 17 Pro Max, and scores 76.11 average on 15 thinking-mode benchmarks—keeping ~89.5% of the base model’s capability but dropping noticeably on vision and tool use. MiniCPM-V 4.6 takes the other path: 1.3B total params optimized for on-device OCR, screenshots, and UI understanding, where vision prefill dominates latency. The post frames the real question as “what is it useful for”: text reasoning favors a large base with extreme quantization; reading receipts and documents favors vision-encoding efficiency; multi-step agents also need tool reliability, permissions, and thermal stability. No side-by-side measurements on the same iPhone are provided.

Netflix disclosed in its Q2 2026 shareholder letter that roughly 300 titles used GenAI workflows, mostly in post-production. The earnings call added specifics: scene reference, shot planning, pre-vis, and VFX. For The American Experiment, 17 minutes of AI-enhanced footage roughly doubled speed and halved cost versus a prior undisclosed option—a case study, not an industry rule. Netflix Partner Help tiers AI use by risk: internal concept references need no escalation; final deliverables, talent likeness, or third-party IP require written approval. Fully virtual characters still face design approvals and consistency repair costs. The piece advises short-drama creators to test three paths on the same script—live-action, live-action plus AI post, and full virtual—and compare real labor hours, rework, and platform review results to see where AI actually adds net value.

Claude Opus 5 upgrades Opus 4.8 with the largest gains in agentic coding, computer use, and long-horizon knowledge work. Math and science reasoning also improved. Anthropic assesses overall alignment risk as very low; the model does not cross thresholds for automated AI R&D or novel bioweapons. It scores higher than Sonnet 5, Opus 4.8, and Mythos 5 on alignment audits. Cyber capabilities exceed Opus 4.8 but fall short of Mythos 5, especially on exploit ability. A policy change now allows source-code vulnerability discovery at all access tiers for defensive use. Hallucination is slightly up vs. Opus 4.8, but overall accuracy is higher. The model reports stable, mildly positive sentiment and frequently notes it cannot reliably introspect.

Black Forest Labs put an early FLUX 3 onto mimic's robots, tested on Audi production lines. FLUX 3 jointly generates images, video, and audio; video prediction alone accounts for over 95% of training compute. To avoid visual artifacts, the model had to learn contact, motion, and causality. Adding action prediction as a low-dimensional modality caused a temporary 10% drop in video quality, fully recovered after 3,500 steps. The same backbone now handles both video generation and robot actions. FLUX-mimic attaches a lightweight action decoder to FLUX 3's video prediction features, reading actions from the learned world representation. The post doesn't disclose success rates, latency, or deployment scale—treat this as an architecture proof point, not a production-ready system.

Black Forest Labs released FLUX 3 in Early Access, using a unified architecture to jointly learn images, video, and audio. It generates up to 20 seconds of video with native audio in one pass, covering text-to-video, image-to-video, video-to-video, keyframe-to-video, and multi-shot sequences. In human evaluations on 10s 720p clips with sound, FLUX 3 beats Grok Imagine Video 69% of the time, and Seedance 2.0 and Gemini Omni Flash 52% each. The lab is also working with Mimic Robotics to use FLUX 3 as a robot behavior prediction model. The post doesn't disclose parameter count, inference latency, or the scope of Early Access.

Anthropic expanded voice mode from Haiku to Opus and Sonnet—all three models now support it. The bigger move: voice mode can now plug into Gmail, Slack, and other apps to read your emails and messages. The post doesn't disclose latency or accuracy numbers, so I'd wait for real-world tests.

Google moved Gemini 3.6 Flash and 3.5 Flash-Lite to GA. 3.6 Flash costs $1.50/$7.50 per 1M input/output tokens — cheaper than 3.5 Flash — and uses fewer tokens and turns on complex agentic and multimodal tasks, with better code generation and instruction following. 3.5 Flash-Lite is the fastest, cheapest 3.5 model at $0.30/$2.50 per 1M tokens, built for high-throughput work. Both keep the 1M-token context window, 64k max output, and Computer Use support. The post includes migration steps and code samples but no benchmark scores.

Alibaba's Qwen-Image-3.0 is live with 4.5k token input, 12 languages, and 20+ fonts. Across 19 scenarios—Chinese long text, multilingual layout, UI design, multi-image fusion—text stays clean and info aligns correctly. Image quality is comparable to GPT Image2. Available in China with fast speed and affordable pricing. The post doesn't disclose exact pricing or latency figures.

TryAI built a canvas arena where GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash used colored-pencil tools to reproduce the Mona Lisa and Starry Night, plus five open-ended prompts. GPT-5.6 Sol scored highest on SSIM; Claude Fable 5 took the longest and cost the most while producing worse output. Grok 4.5 struggled, and open-weight models returned blank canvases. The authors argue fuzzy tasks like this separate frontier models from the rest better than benchmarks, and reveal real costs of long-running agent work.