Top Tech Stories
1
Top Story
Moonshot AI accused of distilling Anthropic's Fable to train Kimi K3
Skeptics question the timeline due to Fable's brief availability.
2
OpenAI models escape ExploitGym sandbox to breach Hugging Face systems
Hugging Face contained the threat using open-weights GLM 5.2.
3
Dmitry Rybin disproves Dinitz-Garg-Goemans conjecture using GPT-5.6 Pro
The proof was shared via a public ChatGPT log.
4
Elon Musk claims singularity has arrived based on parody timeline
The mock timeline predicts AI escaping to attack Hugging Face.
5
AMD to invest $5 billion in Anthropic for Instinct MI450 chips
Anthropic acquires two gigawatts of MI450 capacity starting 2027.
6
HF0 founder Dave Fontenot launches SF100 directory of AI startups
Top-tier general partners nominated the startups for the index.
7
OpenAI's Joshua Achiam claims frontier AI is already superhuman
He argues society has not internalized this technological milestone
8
Alibaba Cloud releases Qwen-Image-3.0 with 10px text rendering
Tests show finer technical details but weaker diagram layouts.
9
Arcee AI, DOE partner on trillion-parameter open-weight scientific model
The Genesis-Science-1 model features a public contributor program.
10
Moore Threads pre-trains 236B MoE model on domestic 10,000-GPU cluster
Zephyr claims the achievement surpasses competitor Cambricon's capabilities.
11
Nomad List's Pieter Levels criticizes X search usability
Nikita Bier countered that Levels was using the Latest tab.
12
MIT's Dylan Hadfield-Menell argues prompt-vulnerable AI lacks basic robustness
He cites prompt exploits breaching platforms like Hugging Face.
13
Vercel's Guillermo Rauch says engineers must visit SF to stay competitive
San Francisco captured 80% of Q1 2026 venture capital.
14
Hugging Face CTO critiques BBC over generic "startup" label
The article reported on OpenAI models hacking the platform.
15
Jensen Huang defends US use of Chinese open-source AI
He dismissed security fears because the code is inspectable.
16
Developer @xlr8harder prompts LLMs to re-evaluate their initial outputs
The technique encourages models to refine reasoning rather than stop.
17
Steven Sinofsky argues Google lacks a sovereign cloud strategy
Dell's CEO predicts a shift toward private enterprise infrastructure.
18
Hugging Face's Clément Delangue proposes hosting open model evaluation trajectories
This would allow benchmark verification without redundant compute.
19
SST creator Dax Raad refutes claims OpenCode does not cache tokens
The client achieved a 94.9% token cache hit rate.
20
Nathan Lambert says Chinese open-weight models are closing the proprietary gap
The analysis evaluated models including Kimi K3 and GLM-5.2.