Samwise Tech/AI/Robotics Newsletter
Wednesday, July 1, 2026
Google's Gemini Omni Flash hits the API, turning enterprise video production into a conversation
Google launched Gemini Omni Flash to its API on June 30, making its conversational video-generation model available to enterprise developers after debuting at I/O 2026. Unlike text-to-video tools, Omni Flash accepts text, images, and video as inputs and supports multi-turn editing — feed it a photo of a specific object and it places that item accurately into a generated scene. The model costs $0.10 per second of output at 720p, putting a ten-second clip at roughly a dollar. Every Omni clip carries Google's SynthID watermark, and the launch is paired with a new AI Content Detection API that flags AI-generated media across all major providers.
Sources: VentureBeat Share ↗ ✉︎ Email 💬 Text
Meituan open-sources LongCat-2.0, a 1.6T near-frontier coding model trained entirely on Chinese chips
Meituan has open-sourced LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts coding model that spent two months as the anonymous top performer on OpenRouter under the codename "Owl Alpha." The model activates 33 to 56 billion parameters per token, supports a 1 million-token context window, and scored 59.5 on SWE-bench Pro — near the frontier of open-weight coding models. What makes the release notable beyond its performance: the entire training run was completed on a 50,000-card domestic compute cluster using Chinese-manufactured chips, with no reliance on restricted Nvidia H100s or AMD MI300Xs. Weights and code are available under a permissive MIT license on Hugging Face.
Sources: VentureBeat Share ↗ ✉︎ Email 💬 Text
Supreme Court rules geofence warrants require probable cause in major privacy win
The U.S. Supreme Court ruled 6-3 on June 29 that police need a probable-cause warrant before demanding historical location data from tech companies through geofence requests. Writing for the majority, Justice Elena Kagan held that people retain a reasonable expectation of privacy in their cell-phone location data, bringing geofence searches under Fourth Amendment protections. The ruling does not ban geofence warrants outright — law enforcement can still obtain them — but mandates the same warrant standard applied to traditional searches. Justices Alito, Thomas, and Barrett dissented, with Alito calling the decision an "irresponsible escapade" that destabilizes Fourth Amendment jurisprudence.
Sources: TechCrunch Share ↗ ✉︎ Email 💬 Text
Morgan Stanley cut its riskiest reconciliation job in half by making its AI agents less autonomous
Morgan Stanley has cut its highest-risk profit-and-loss reconciliation workload in half using an agentic system called FIXR — and the key to making it work was reducing autonomy rather than increasing it. The bank deliberately keeps humans in the loop for every reconciliation decision; when controllers flag a break type repeatedly, the system converts that pattern into a fixed rule instead of relying on model judgment. "It's much more like a co-worker than a copilot," said Managing Director Todd Johnson. The approach reduces token consumption, improves audit-trail predictability, and has made the case inside Morgan Stanley: tighter human-AI coupling, not full automation, is the right path for high-stakes financial workflows.
Sources: VentureBeat Share ↗ ✉︎ Email 💬 Text
DeepSeek and Peking University open-source DSpark, speeding LLM inference up to 85%
DeepSeek and Peking University jointly open-sourced DSpark, a speculative decoding framework that accelerates LLM inference 60 to 85 percent without changing model outputs. DSpark's semi-autoregressive design targets suffix decay — a known bottleneck in sequential token generation — and uses a confidence-scheduled validation method to maintain accuracy. Benchmarks on DeepSeek-V4 Flash show per-user generation speeds up 60 to 85 percent over the MTP-1 baseline; V4 Pro models gained 57 to 78 percent. The release also includes DeepSpec, a full-stack training toolchain designed to let smaller developers build low-latency inference services without the infrastructure budgets of major AI labs.
Sources: VentureBeat Share ↗ ✉︎ Email 💬 Text
Pocket raises $11M on 130,000 units sold for its AI recording puck that needs no subscription
Pocket, maker of a $129 credit-card-sized recording puck that attaches to the back of a phone, has raised $11 million led by Accel with participation from Y Combinator, ElevenLabs CEO Mati Staniszewski, Vercel CEO Guillermo Rauch, and Opendoor CEO Kaz Nejatian. The company has shipped more than 130,000 units since launch and offers unlimited recordings, transcriptions, and to-do lists with no subscription. The raise validates the thesis that physical recording hardware can build a durable business on a one-time device price rather than SaaS recurring revenue. Pocket will use the funding to expand hardware form factors and build out enterprise capabilities.
Sources: TechCrunch Share ↗ ✉︎ Email 💬 Text
X Square Robot reaches $2.8B valuation across four consecutive funding rounds backed by Xiaomi and IDG
Shenzhen-based X Square Robot Technology has closed four consecutive funding rounds, culminating in a Series C, pushing its valuation above $2.8 billion. Backers include IDG, HongShan, and Xiaomi, which have supported the company across multiple rounds. X Square builds end-to-end embodied AI systems anchored by WALL-B, a foundation model based on its World Unified Model architecture that trains perception, language, action, and physical prediction in a single unified network rather than separate modular components. The company has partnered with 58.com to run AI-powered cleaning services in Shenzhen and Beijing homes, and its X Family Member Program places robots with households for month-long companion trials.
Sources: The Robot Report Share ↗ ✉︎ Email 💬 Text
Tech Pulse
Top Frontier Models (SWE-bench Verified): Claude Mythos 5 (95.5%) | Claude Fable 5 (95.0%) | Claude Opus 4.8 (88.6%)
Top Open Source Models (SWE-bench Verified): DeepSeek-V4-Pro-Max (80.6%) | MiniMax M3 (80.5%) | Qwen3.7 Max (80.4%)
Top Small Models 15–50B (SWE-bench): Gemma 4 31B (80.0%) | Mistral Medium 3.5 (77.6%) | Qwen 3.6-27B (77.2%)
Top Edge Models 0–15B: Phi-4 (14B) | Gemma 3 (9B) | Qwen3.5 (7B) (SWE-bench n/a at this tier)
AI Leaders (market cap): NVIDIA $5.23T | Alphabet $4.63T | Microsoft $3.11T
Robotics Leaders (market cap): Intuitive Surgical ~$150B | ABB $47B | Fanuc $30B
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.