Author: Amitabh Sarkar

Published: August 15, 2026 Anthropic is in talks to acquire Israeli AI startup Decart AI for approximately $6 billion, Bloomberg reported August 13, 2026, citing people familiar with the discussions. The deal would be Anthropic’s largest known acquisition to date. Decart builds software that reduces the cost of training and running AI models by making chips work more efficiently, and the talks have not concluded — the price and terms could still change. Decart raised a funding round in May 2026 that valued the company at nearly $4 billion, up from $3.1 billion in August 2025. A $6 billion price…

Google published HEIR (Homomorphic Encryption Intermediate Representation) on August 15, 2026 — an open-source compiler that converts pre-trained AI models to run on encrypted data, allowing cloud servers to generate predictions without ever seeing the underlying plaintext. HEIR is part of Google’s Private Computing Toolkit and is backed by hardware accelerator partners Belfort, Niobium, Cornami, and Optalysys, who are developing chips to reduce the historically high compute cost of homomorphic encryption. What Is HEIR and What Does It Do? HEIR is a compiler toolchain that converts existing AI models — built on standard frameworks — to operate on homomorphically encrypted…

Alibaba’s Qwen team released Qwen3.8-27B on August 14, 2026 — a 27.78-billion-parameter dense model licensed under Apache 2.0 that runs on 17GB of VRAM at 4-bit quantization, making it the strongest self-hostable open-weight AI agent in its size class at release. The model supports text, image, and video inputs, carries a native 262,144-token context window (extendable to 1 million tokens via YaRN), and generates 131 tokens per second on a single RTX 5090. Unlike its companion model Qwen3.8-Max — a 2.4-trillion-parameter mixture-of-experts model available via cloud API — the 27B is built for private, on-premise deployment with zero marginal inference…

Z.ai released GLM-5.3 on August 14, 2026, claiming the strongest open-weights coding model it has measured and disclosing that the model developed multi-step exploit-chain reasoning its training never planned for. On August 26, 2026, Z.ai fulfilled its promised open-weights release by shipping GLM-5.3-Flash — a lighter, natively multimodal MoE branch — under the MIT license on Hugging Face. GLM-5.3 is post-trained on the same 743B-parameter base as GLM-5.2, so every reported gain comes from post-training rather than a larger model. Z.ai reports a 50% improvement in coding capability over GLM-5.2 in internal evaluations, first place among open-source models on Terminal…

Wix launched Symphony on August 11, 2026, a standalone multi-agent AI platform that learns a small business’s values, practices, and goals, then automates its workflows and reports back in a daily stand-up briefing. Symphony runs separately from the Wix website builder and targets small and medium-sized businesses rather than enterprises. A single orchestrator agent named Maestro sits at the centre of Symphony. Maestro identifies what needs attention and coordinates specialized agents per activity, while a second class of accuracy-checking agents validates each output before the business owner sees it. Wix built the industry-specific behaviour on data drawn from what the…

DeepSeek released Harness v0.1 on August 13, 2026, an MIT-licensed open-source AI agent runtime designed to sit between a large language model and the tools, files, and environments an agent needs to complete multi-step tasks. The release, codenamed “Black Whale,” is an explicit developer preview that frames Harness as a modular, self-hostable alternative to closed coding-agent environments like Claude Code. The launch landed on the same day as the general availability of DeepSeek V4-Pro on the API. Harness drew 608 points and 261 comments on Hacker News by August 14, making it one of the top developer discussions of the…

Google launched Gemini 3.7 Flash on August 13, 2026, scoring 43.6% on the FrontierCode 1.1 Main benchmark — above Claude Sonnet 5’s 42.7% — and priced it at $0.75 per million input tokens through December 31, 2026, half the rate of its predecessor Gemini 3.6 Flash. The model is live in Gemini Spark and available via the Google AI API. The launch makes Gemini 3.7 Flash the most cost-performance-competitive model Google has shipped for coding and AI agent workloads in the Flash tier, arriving three weeks after Gemini 3.6 Flash in a cadence that has now produced three Flash-tier releases…

Published: August 13, 2026 Lovable raised $400 million in a Series C round at a $13.3 billion valuation, the Stockholm-based AI app builder announced August 12, 2026. Menlo Ventures led the round, with Scaleup Europe Fund (EQT) co-leading, and the valuation roughly doubles the figure investors assigned the company in its December 2025 Series B round. Lovable lets users describe software in plain language and receive a working app, a category the company calls “vibe coding.” Since its November 2024 launch, users have created more than 60 million projects on the platform, and apps built on Lovable now receive more…

Published: August 13, 2026 Researchers at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems extracted hidden chain-of-thought reasoning from OpenAI, Anthropic, and Google models by exploiting a shared encryption key flaw, according to a paper submitted August 10, 2026. The attack recovered raw reasoning from GPT-5.5, Claude Haiku 4.5, and Gemini before all three providers patched the flaw following coordinated disclosure. OpenAI, Anthropic, and Google return chain-of-thought reasoning to API clients as an encrypted block that clients cannot read directly. The researchers found that every model within the same family shared one encryption key, turning the…

Published: August 12, 2026 Anthropic began embedding machine-readable watermarks into all text produced by Claude models released on or after August 2, 2026, making every output traceable after copy-paste, driven by EU AI Act Article 50 transparency requirements and applied globally. The watermark is woven into statistical patterns of word selection and phrasing — invisible to human readers but detectable by machines. It signals that Claude processed a piece of text, not that Claude authored it, meaning businesses that use Claude to edit or summarise human-written documents will still produce watermarked output. Anthropic has confirmed detection tools are coming, though…