AI Updates on 2025-08-13

OpenAI releases updates to GPT-5 with new control options allowing users to choose between "Auto", "Fast", and "Thinking" modes, increased rate limits to 3,000 messages/week for GPT-5 Thinking, and 196k token context limit @sama
Google introduces personalization features for Gemini app, allowing the model to learn from past conversations and offering temporary chat mode for sensitive conversations @GeminiApp
Anthropic releases Claude Code with new "Opus plan mode" that uses Claude Opus 4.1 for planning and Claude Sonnet 4 for execution @_catwu
Perplexity launches Comet desktop application for all US-based Pro users, featuring Max Assistant mode for Max subscribers with advanced reasoning capabilities @perplexity_ai

Anthropic's focus on developers is making them the preferred choice across tech companies, with one scaleup founder switching entire team to Claude Enterprise subscriptions due to GPT-5 hallucination issues @GergelyOrosz
AI evaluation test suites now add token costs as a new consideration for CI/CD pipelines, with one startup CTO reporting $50 per test suite run @GergelyOrosz
NVIDIA emerges as the leading open model ecosystem lab in the US over the past 6 months, according to industry analysis @natolambert
Research reveals 41% of YC-backed AI startups are building tools workers don't want, representing a $50B market misalignment @FounderCoHo
Commonwealth Bank, Australia's biggest bank, announces new partnership with OpenAI @gdb

François Chollet warns that generative AI acts as "informational pollutant" and "cognitive smog" that corrupts internet content, transforming human expression into "uniform, gray slurry of derivative outputs" @fchollet
AI Now Institute highlights concerns about Big Tech and federal government alliance positioning major AI companies as "too big to fail" @AINowInstitute
Anthropic shares detailed post on their Safeguards team's approach to identifying potential model misuse and building defenses, covering policy development, training, testing, and real-time monitoring @AnthropicAI
Reid Hoffman discusses Taiwan's use of AI-facilitated "alignment assemblies" to combat deepfake scams and build democratic consensus, demonstrating how AI can strengthen rather than undermine democratic processes @reidhoffman

Perplexity Finance expands to Indian markets, offering synthesis of Indian markets news, live stock prices for BSE & NSE equities, and natural-language stock screening features @AravSrinivas
Microsoft Research releases RetroChimera on Azure AI Foundry for predicting synthesis routes to drug-like molecules, advancing AI applications in drug discovery @MSFTResearch
Stability AI and NVIDIA collaborate to deliver 1.8x faster Stable Diffusion 3.5 performance through NIM microservice with streamlined enterprise deployment @StabilityAI
Paul Graham shares example of using ChatGPT to help respond to anti-vaccine conspiracy theories, demonstrating practical family communication applications @paulg
PyTorch releases ExecuTorch 0.7 bringing KleidiAI acceleration to billions of Arm devices, including 3-5 year old phones and Raspberry Pi 5 for on-device AI @PyTorch

GPT-5 (Thinking medium) now far exceeds medical professionals on medical reasoning benchmarks, while GPT-4o was previously below their level @emollick
Researchers extract base model from OpenAI's GPT-OSS, revealing strong underlying capabilities beneath the reasoning-only interface and releasing gpt-oss-20b-base @jxmnop
Andrew Curran reports GPT-5-thinking shows exceptional performance at interpreting hidden meaning and intent in short stories, calling it "the best I've ever seen at this" @AndrewCurran_
Aidan McLaughlin highlights impressive cognitive capabilities in AI models combining spatial IQ, long-horizon coherence, and aesthetic judgment using mcbench evaluation @aidan_mclau
Hugging Face releases new TRL version with native supervised fine-tuning support for vision language models, multimodal GRPO, and MPO capabilities @mervenoyann
Chinese models dominate open model performance rankings across most benchmarks, with top half occupied by Chinese models and bottom half by everyone else @natolambert