WAIC 2026 Opens as Embodied AI Takes Center Stage

Headlines

WAIC 2026 Opens, Embodied AI Moves From Demo to Deployment

The 2026 World Artificial Intelligence Conference (WAIC 2026) opened in Shanghai on July 17, under the theme "Intelligent Partners, Shared Future."

On-the-ground coverage from multiple tech outlets points to the same signal: embodied AI is the headline theme this year. TMTPost captured it as "embodied intelligence takes center stage for the first time," with the telling shift — "last year's demos, this year's sales" — as the industry's focus moves from tech showcases to actual shipments.

Yin Qi, chairman of StepFun and Qianli Technology, gave a keynote titled When Agents Enter the Physical World. Reflecting on 15 years in AI, he argued that model capability is crossing a critical threshold in 2026, and that AI entrepreneurship has gone from a niche to a global consensus.

Source: 36Kr · TMTPost

Apple Sues OpenAI Over Trade Secrets, at the Worst Possible Time

Apple filed a trade-secrets lawsuit against OpenAI last Friday, and it isn't messing around. The complaint alleges a pattern of misconduct reaching all the way up to OpenAI's chief hardware officer, and claims that more than 400 former Apple employees now work at the company.

Multiple outlets note the timing couldn't be worse — it lands right as OpenAI pushes toward an IPO. Some commentators argue many of the allegations describe how things are commonly done in the industry, and that Apple's move reads more like a public show of pressure.

Source: TechCrunch · The Verge

DeepSeek V4 Flash: A Million-Token Context on a Single Consumer GPU

After DeepSeek V4 Flash's weights dropped, community local-inference benchmarks rolled in. Two data points are worth keeping:

One developer ran DeepSeek V4 Flash (Unsloth's Q8 quant) on a single RTX 5090 via llama.cpp with a 1-million-token context, noting that recent llama.cpp changes made it "much more usable."

A second comparison is more striking: the same 89-task Terminal-Bench 2.1 suite scored 54% on a 128GB M5 Max MacBook (an aggressively quantized 80.8 GiB GGUF), versus 52% on 2× DGX Spark (native FP8/FP4 with speculative decoding) — one laptop matching, even edging out, two dedicated boxes.

Read together, both point the same way: open weights plus consumer hardware are turning "run a large model locally" from an enthusiast experiment into a reproducible baseline.

Source: LocalLLaMA (5090 / 1M context) · LocalLLaMA (MacBook vs DGX Spark)

Technical & Insight

Kimi K3 Keeps Topping Community Benchmarks; the Open-Source Push Accelerates

Following last issue's spec reveal, third-party and community results for Kimi K3 landed this week: on Simple Bench, Kimi K3 (max) beat Claude Sonnet 5; filtered for science queries on Text Arena, K3 sits at the top.

Developer Simon Willison also ran K3 through his signature "pelican riding a bicycle" SVG benchmark and wrote up an analysis.

A recurring narrative in the community: Kimi K3, DeepSeek V4 Flash, plus the imminent GLM and Minimax releases — open models are accelerating as a group. Some argue this collective momentum is exactly where open source's strength lies.

A caveat: leading one benchmark isn't the same as being best overall. Last issue, Moonshot's own blog stated plainly that K3 "still trails the strongest closed models" — the same model can sit in very different places across dimensions, and any single number is easy to over-read.

Source: LocalLLaMA (Simple Bench) · Simon Willison

Product & Governance

Databricks Hits a $188B Valuation

Databricks closed a new round at a $188 billion valuation, extending its run as "AI's favorite second act."

In recent years Databricks has remade itself into an AI company, and has published research on the cost advantages of open-weight models for coding tasks.

Source: TechCrunch

TikTok Tests an AI Likeness Detection Tool

TikTok has begun testing an opt-in tool that scans for AI-generated likenesses and lets creators report them to the platform.

Spotted by social-media consultant Matt Navarra and confirmed to The Verge by a TikTok US spokesperson, the tool is initially being tested with "some" US creators. As AI face-swaps and digital doubles grow more common, it's an attempt to turn likeness protection into a platform-side product feature.

Source: The Verge