Google Announces Real-Time Speech Translation with Gemini 3.5 Live Translate
On June 9, Google DeepMind released "Gemini 3.5 Live Translate," a real-time speech-to-speech translation model. It automatically detects more than 70 languages and generates translated speech while preserving the speaker's tone and emotion. Google Translate started as a machine learning experiment 20 years ago and now translates over a trillion words every month; what sets this model apart is that it goes beyond text translation to reproduce the very "texture of the voice." It is expected to be put to use in international conferences and customer support.
Amazon Employees Mock Internal AI as "Sloppenheimer"
According to a report by 404 Media, Amazon employees have been mocking the company's AI-generated content as "Sloppenheimer" on internal Slack. The coinage blends "sloppy" and "Oppenheimer," and it lays bare internal frustration with the low-quality code and documents that AI produces. The story also drew significant attention in the developer community, picking up 90 points and 43 comments on Hacker News. The incident is once again calling into question the balance between the speed of AI adoption and quality control.
Apple's AI Strategy: Partnering with Google for Siri, but No Gemini in Its Own Models
Apple's AI strategy is showing two faces. Craig Federighi offered a detailed account of the company's partnership with Google for the Siri AI in iOS 27. At the same time, it was also reported that Apple's in-house AI models contain no Google Gemini components whatsoever. The apparent policy is to leverage partners' technology to improve the user experience while keeping the core models self-reliant. How Apple balances building its own AI as a platform company against reliance on external APIs will be worth watching.
Anthropic × Blackstone JV Acquires Fractional AI
The Blackstone-Anthropic joint venture has announced the acquisition of Fractional AI, an AI consulting firm. The move appears to be part of a strategy to bring Anthropic's Claude to enterprise use, and it is accelerating the trend of AI vendors strengthening their direct customer relationships. As vertical integration advances across the AI industry, the picture of model developers pushing into the services layer themselves is coming into sharper focus.
The Open-Model Camp Stays Busy Too: Cohere North Mini Code 1.0 and Unsloth Gemma 4 QAT
New models have also landed one after another in the open-weight local LLM space.
Cohere has released the official version of "North Mini Code 1.0," a coding-focused 30B A3B (3B active parameters) model. On Artificial Analysis scores it falls short of Qwen 3.6 35B overall, but it remains competitive on coding metrics, where it is said to beat Gemma 4 26B.
Meanwhile, Unsloth has released a full lineup of GGUF models with QAT (Quantization-Aware Training) and MTP (Multi-Token Prediction) support for Google's Gemma 4 series (12B/26B/31B/E2B/E4B), enabling fast inference in local environments.
Accepted at ICML 2026: Predicting and Controlling Hallucinations with ntkMirror
"Predictable Compression Failures," a paper accepted at ICML 2026, is attracting attention. The paper provides a theoretical analysis of the "order variance" phenomenon in evidence-based QA tasks, where the order in which evidence is presented changes the answer probability, and designs a gate that makes the model abstain when the available information is insufficient.
The implementation, ntkMirror, can be applied to open-weight models without any additional training. Validated on models including Qwen2.5-7B and Gemma E4B, it achieved hallucination rates of 0.0–0.7% while abstaining on roughly 24% of questions — a result that looks poised to improve the reliability of local LLMs.
---