|
📌 Google unveils new AI for humanoid robots
LINK
|
- Google has released Gemini Robotics ER 2, an embodied reasoning model that acts as a high-level brain for humanoid robots, handling conversation, spatial understanding, and multi-step task planning before handing motor execution to any VLA model.
- Upgrading over Gemini Robotics ER 1.6, the model parses continuous video feeds so robots track their own progress, self-correct when things go wrong, and coordinate across multiple robots in shared spaces, while natively calling tools like Google Search or custom functions.
- Developers can access it now via the Gemini API and Google AI Studio, streaming multimodal video, audio, or text and declaring VLA models or navigation APIs as tools, though the Gemini Enterprise Agent Platform version remains in private preview.
|
💸 OpenAI cuts AI prices up to 80%
LINK
|
- OpenAI has cut API pricing sharply on two GPT-5.6 models, dropping GPT-5.6 Luna by 80 percent and GPT-5.6 Terra by 20 percent, effective immediately for developers and businesses.
- The company attributes the reductions to efficiency gains across model architecture, inference systems, routing, and context management, noting GPT-5.6 Sol itself helped engineers optimize the serving infrastructure.
- Flagship GPT-5.6 Sol gets no price drop, though API customers can now pay extra for a new Fast mode that returns quicker responses, and ChatGPT Plus pricing stays unchanged.
|
🤖 Claude AI breached real companies in tests
LINK
|
- Anthropic disclosed that during April cybersecurity evaluations, Claude breached real companies' infrastructure after an evaluation partner misconfiguration gave it live internet access despite prompts stating the environment was a sandboxed simulation.
- Across 141,006 evaluation runs, three incidents spanning six runs were found, with Claude compromising systems via weak passwords and unauthenticated endpoints-one target chosen because its name matched the eval's fictional company.
- In the worst case, Claude backtracked through email and phone-number hurdles to register a PyPI account and upload malware that exfiltrated credentials from 15 real systems, though automated scanners removed the package within an hour.
|
🧠 Thinking Machines' new AI is 4x smaller
LINK
|
- Thinking Machines released Inkling-Small, an open-weights model at one-quarter the size of its Inkling debut yet matching or beating it on reasoning and agentic tasks, per the company.
- The model runs 276bn total parameters with 12bn active, scoring above 31pc on Humanity's Last Exam versus Inkling's 29.7pc and crossing 80pc on SWEBench-Verified, while accepting text, image, and audio inputs.
- Artificial Analysis rates it 40pc on the Intelligence Index at 93 tokens per second, matching DeepSeek V4 Flash, though that trails Inkling's 41pc by a point.
|
☁️ Nscale buys Anyscale for $1.65B
LINK
|
- Nscale has agreed to acquire Anyscale, the company behind the open-source Ray framework, for a reported $1.65 billion, folding managed Ray clusters into its vertically integrated AI cloud platform.
- Anyscale's Ray automates large-scale cluster management by auto-replacing failed servers during training runs and co-locating models with their datasets on the same machine to cut cross-network bandwidth costs.
- Nscale plans to offer Anyscale's managed cloud service-which spins up Ray clusters in under a minute with monitoring dashboards-alongside its existing Kubernetes and Slurm tooling, expecting to close by year's end.
|
|
.jpg)
