From Brain Waves to Words
Version 2 of brain2qwerty, one of Meta’s open-source neuroscience projects targeting communication via brain waves, has been announced. The model was trained on 22,000 sentences typed by 9 volunteers using an MEG device over 10 hours, and achieved 61% word accuracy. That’s a significant step forward compared to the 8% accuracy rate of other non-invasive models.
New Models, New Restrictions
This week we heard that OpenAI’s new GPT-5.6 model has also been restricted by the United States. Anthropic’s banned model Fable is set to return tomorrow. They’ve significantly increased the safety guardrail layers: for coding-related queries, for example, the model will automatically switch to Opus 4.8. Meanwhile, Anthropic also announced the 5th version of the Sonnet model, which drew criticism for being more expensive than Opus 4.8 despite underperforming it at launch.
TechCrunch → | Claude AI (X) →
From the Model Race to the Chip Race
Alongside new models, how and with what hardware models are trained continues to dominate the conversation. Meituan, China’s largest food delivery company, announced that their new open-source model called LongCat was trained entirely on locally produced Chinese chips. OpenAI, meanwhile, unveiled its first custom chip called Jalapeño, developed in partnership with Broadcom, as part of its effort to reduce dependence on Nvidia.
Bonus: Worth Your Screen Time
This video on the engineering philosophy behind systems that reliably process trillions of transactions is, in our view, worth adding to your watch list.
⚡️ Quick Bits
• Amazon joined the forward deployed engineer (FDE) trend. TechCrunch →
• Apple acquired Play, a prototyping app. 9to5Mac →
• Openclaw announced its iOS and Android apps. TechCrunch →
• arXiv declared its independence. Science →
📩 Liked this newsletter? You can forward it to your friends
if you are feeling nostalgic, though it’s always better if
they just subscribe on their own.
