AI NEWS DAILY

OpenAI released GPT-5.6, their latest model iteration with significant capability improvements across reasoning, coding, and multimodal tasks.

AI Futures Project publishes 'Plan A' — a scenario for delaying superintelligence until 2040 through international cooperation, total research transparency, and mutually assured compute destruction.

EPFL researchers developed AI-generated videos designed to maximally activate a target brain region, using fMRI data to optimize visual stimuli.

OpenAI launched ChatGPT Work, an enterprise version of ChatGPT designed for professional workflows and organizational use.

Pangram study finds AI-generated content is pervasive across social media, with LinkedIn as the worst offender.

Ello blog post details building a real-time AI tutor optimized for 5-year-old learners with 1000ms latency.

Developer reports chronic 'LLM burnout' from hours of daily interaction with AI assistants, describing a shift from writing code to designing, prompting, and reviewing AI-generated code.

Cognition launched SWE-1.7, reaching frontier-level intelligence (near GPT-5.5 and Opus) at much lower cost, trained from a Kimi K2.7 base with extensive RL post-training.

Brown University professor suspected AI cheating on take-home exams; in-person final scores dropped 50%.

Side-by-side comparison of Grok 4.5, GPT-5.5, and Claude building identical applications.

Databricks benchmarks AI coding agents against their own multi-million-line production codebase.

Hacker News discussion seeking alternative tech news platforms with less AI content and more human hacking.

Noma Labs discovered a prompt injection vulnerability in GitHub's Agentic Workflows that lets attackers silently extract data from private repos via crafted public issues.

Slopfix charges $10,000/week to refactor AI-generated codebases that have become unmaintainable — analyzing for free, then cutting code bloat with a committed reduction target.

OpenAI announces public launch of GPT-5.6 Sol alongside Terra and Luna, possibly satirical naming.

Developer investigates YC CEO's claim of shipping 37K lines of AI-generated code daily and looks under the hood.

GLM-5.2 by zai-org is a multilingual text-generation model with MoE-DSA architecture supporting English, Chinese, and Arabic.

DeepSeek-V4-Pro by deepseek-ai is a conversational text-generation model with 8-bit/FP8 quantization support.

DeepSeek-R1 is a reasoning-focused text-generation model using the deepseek_v3 architecture with FP8 support.

Llama-3.1-8B-Instruct by Meta is an 8B-parameter instruction-tuned model supporting 8 languages.

FLUX.1-dev by Black Forest Labs is a text-to-image generation model compatible with the diffusers library.

Gemma-4-12B-coder-fable5-composer2.5-v1-GGUF is a GGUF-quantized coding model fine-tuned from Google's Gemma-4-12B.

Meta-Llama-3-8B is Meta's original 8B-parameter base language model from the Llama-3 family.

Llama-2-7b-chat-hf is Meta's 7B-parameter chat-tuned model from the Llama-2 generation.

Meta-Llama-3-8B-Instruct is Meta's instruction-tuned 8B model from the Llama-3 family with Azure deployment support.

BLOOM by BigScience is a 176B-parameter multilingual model supporting 46 languages across 13 families.

gpt-oss-120b by OpenAI is a 120B-parameter open-source text-generation model with MXFP4 quantization.

gpt-oss-20b by OpenAI is a 20B-parameter open-source text-generation model with MXFP4 quantization support.

Phi-2 by Microsoft is a compact 2.7B-parameter model trained on synthetic 'textbook-quality' data.

Stable Diffusion XL 1.0 by Stability AI is a high-resolution text-to-image model with 140K+ downloads.

Mistral-7B-Instruct-v0.2 by Mistral AI is a 7B-parameter instruction-tuned model with Apache-2.0 licensing.

Mistral-7B-v0.1 by Mistral AI is the original 7B-parameter pretrained base model from Mistral.

DeepSeek-V3 is a conversational text-generation model using the deepseek_v3 architecture with FP8 support.

Stable Diffusion v1.4 by CompVis is the original 410M-parameter text-to-image diffusion model.

Llama-3.3-70B-Instruct by Meta is a 70B-parameter instruction-tuned model supporting 8 languages.

Gemma-7B by Google is a 7B-parameter open-weight model with extensive research paper references.