12 Aug 2025

AI News Digest

🤖 AI-curated 7 stories

Today's Summary

OpenAI’s GPT-5 launch has been bumpier than expected, with rollout glitches making the model fumble basic tasks and prompting the return of older versions while fixes roll out. Meanwhile, India’s deep-tech scene just got a boost with a $69M fund for AI, space, and climate startups, and researchers dropped Goedel-Prover-V2 — an open-source theorem-proving model that beats prior records with far fewer parameters. Also in AI news: Profluent claims its AI-designed CRISPR editor outperforms the classic Cas9, Google added a tutor-style “Guided Learning” mode to Gemini, and OpenAI quietly released two open-weight reasoning models that can run on consumer hardware.

Stories

OpenAI’s GPT-5 rollout hits glitches as users report basic errors

OpenAI’s highly anticipated GPT-5 — the newest flagship model powering ChatGPT — has faced a rocky public rollout, with users posting examples of the model making simple math and geography errors and complaining about reduced access to older models. Axios reports the problems stem in part from rollout routing issues that kept some queries from reaching GPT-5’s slower, higher-quality “reasoning” mode; OpenAI CEO Sam Altman acknowledged glitches, restored older models, and said fixes and clearer UI indicators are coming. Why it matters: the stumble underscores how technically and reputationally risky major model launches are, even for top AI firms, and highlights the challenge of balancing speed, access, and quality for large multimodal systems — with implications for customer trust, enterprise adoption, and competitive positioning in the fast-moving generative-AI market.

Indian VC Speciale Invest pledges $69M to back deep‑tech startups including AI

Speciale Invest announced a plan to deploy 6 billion rupees (about $69 million) through 2029 to back seed and pre-seed deep‑tech startups in India, with a focus that explicitly includes artificial intelligence, space tech, climate tech and dual‑use defence. Reuters reports the fund will make initial tickets of 70–100 million rupees across roughly 18–20 companies and reserve over half the capital for follow‑on investments. Why it matters: the commitment adds notable early‑stage capital to India’s AI and deep‑tech ecosystem, signaling growing investor confidence in local generative‑AI and specialized AI startups and potentially accelerating product development, commercialization, and talent growth in the region.

Goedel‑Prover‑V2: New open-source LLM sets state‑of‑the‑art in automated theorem proving

Researchers released Goedel‑Prover‑V2 (arXiv Aug 5, 2025), an open‑source family of language models that advances automated formal theorem proving. The paper describes scaffolded data synthesis, verifier‑guided self‑correction using the Lean compiler, and model averaging; the result is a 32B model that outperforms prior SOTA on MiniF2F and PutnamBench while a much smaller 8B variant matches much larger prior models. This is a notable leap for formal reasoning: higher proof success with smaller models makes formal verification and research‑grade math assistants more accessible to academia and industry, and the team has released code, weights and new benchmarks to accelerate follow‑up work.

Profluent’s OpenCRISPR‑1: AI‑designed gene editor reported in Nature, company opens dataset and tools

Genetic Engineering & Biotechnology News (GEN) reports that Profluent published work (appearing in Nature) describing OpenCRISPR‑1 — an AI‑designed CRISPR‑Cas nuclease the company says matches (or exceeds) SpCas9’s on‑target activity while dramatically reducing off‑target edits and showing lower immunogenicity. The team trained protein language models on a massive CRISPR‑Cas atlas and released the sequences, models and an atlas dataset. If validated broadly, AI‑designed editors could speed therapeutic development and expand molecule design beyond what evolution has produced — but they also raise questions about safety, governance, and careful independent validation.

Read more → GEN (Genetic Engineering & Biotechnology News)

GitHub CEO Thomas Dohmke to step down, plans to found a new startup

Thomas Dohmke, CEO of Microsoft‑owned GitHub, announced he will leave the company to start a new venture and will remain through the end of 2025 to help the transition. The move caps a tenure that included a big push into AI (GitHub’s Copilot products and a reported doubling of AI projects on the platform). Reports say Microsoft will not immediately name a replacement and that parts of GitHub’s leadership will report into Microsoft’s broader developer/AI organization—a shift that underscores tighter corporate integration and highlights intensifying competition in AI-powered coding tools. Industry impact: this is a notable executive change at one of the primary platforms for developer AI tooling, likely to accelerate alignment between GitHub product roadmaps and Microsoft’s CoreAI investments while creating an opening for rivals in the AI coding space.

Google’s Gemini adds Guided Learning — an AI tutor built for real study

Google rolled out a new Guided Learning mode inside the Gemini app that acts like a step-by-step tutor (quizzes, visuals, flashcards and interactive explanations) to help students learn concepts rather than just get answers. The feature ties into Google’s broader education push (including a one-year free AI Pro plan for eligible students) and is designed to integrate images, videos and classroom workflows — a direct response to growing concerns about AI-facilitated cheating and an effort to position AI as a learning aid. This matters because it turns mainstream consumer AI into a practical study tool for classrooms and self-learners, lowering friction for structured learning and prompting educators to rethink how AI is used in instruction. ([theverge.com](https://www.theverge.com/news/732182/google-gemini-ai-guided-learning-education))

OpenAI releases gpt-oss models — powerful open-weight reasoning models for developers

OpenAI published two open-weight reasoning models (gpt-oss-120b and gpt-oss-20b) under an Apache 2.0 license, making advanced reasoning and coding-capable models available for download and local use (Hugging Face and partner platforms). The larger model runs on a single 80GB-class GPU while the 20B variant can run on consumer hardware (~16GB VRAM), and OpenAI emphasizes tool use, chain-of-thought reasoning, and wide deployment options — aiming to democratize access for developers, researchers and smaller organizations. The release could accelerate hands-on experimentation, local development workflows, and new tools/apps that embed stronger reasoning capabilities without relying solely on hosted APIs, while also raising practical questions about safe fine-tuning and misuse. ([techcrunch.com](https://techcrunch.com/2025/08/05/openai-launches-two-open-ai-reasoning-models/))