A short list of videos we think are worth your time. Each one comes with our own summary, the key takeaways and what it means for buyers, so you can decide whether to press play.
A 24-minute tour (the video carries a sponsor segment for Eight Sleep) of projects the creator says Fable 5.2 built with minimal prompting, then reads out vendor benchmark scores and pricing.
Matthew Berman walks through a METR assessment of Anthropic, arguing it shows AI is now the main driver of faster model progress, while noting the access was limited and Anthropic controlled what was published.
Two Minute Papers walks through MIT research in which GPT-6 Astra inspects and edits neural network weights, and spends much of the time correcting how the work has been reported.
Two Minute Papers cites research on AI rebuilding programmes and web pages from behaviour alone, and argues the real constraints are legal, commercial and about control of models.
The channel reads the 264-page system card and a METR report, arguing that Fable 5.2 leads on several benchmarks while common sense and evaluation integrity lag.
Two Minute Papers summarises a DeepMind paper in which AI agents produced machine-checked proofs that the libsignal implementation matches its mathematical specification, within stated limits.
In a conversation with Matthew Berman, Anthropic researcher Jeremy Berman explains neural networks, pre-training, post-training and reinforcement learning using toy examples and small open models.
Video by Matthew Berman on YouTube
7 videos
The ByteOrbital letter
One letter. What mattered in AI and hardware.
Headlines, standfirsts and links to our own coverage, nothing else. We confirm by email first and every message has a one-click unsubscribe.