Magic, the AI startup founded by Eric Steinberger and Sebastian De Ro, published a research update claiming its language-model pretraining is over 10x more compute-efficient than recipes behind leading open-weight base models. It says matching DeepSeek V4 Pro’s pretraining took 50x less compute, costing about $0.5M on Nvidia GB200 (and scaling to about $4M). The result is a new cost-and-benchmark narrative based on its own held-out bits-per-byte loss measurements and a next stated focus on long-horizon reinforcement learning and deployment safety policies, while noting the figures are unverified and no model has shipped.
Pathway developed its brain-inspired BDH (Dragon Hatchling) architecture and ran training on Amazon SageMaker HyperPod for scaling out distributed GPU compute. BDH-CQ reached 29.2% pass@2 on the ARC-AGI-1 benchmark at a stated cost of US $0.0007 per task. The shift updates model reasoning in a recurrent latent state without generating long chain-of-thought traces, changing the cost–accuracy tradeoff and how memory and computation are handled at inference.
The talk “From S3 to GPU in One Copy” explains how SpiralDB’s Vortex columnar format enables training data to be streamed from S3 into GPUs with projection and filter pruning instead of decompressing/reprocessing everything. It reports S3-to-GPU scans at around 30 times faster than Parquet. As a result, only selected columns/segments are read and decompressed for each iteration, reducing movement and decision overhead during ML training.
vLLM explains how speculative decoding in its serving stack uses a draft component to propose multiple tokens and a target-model pass to verify them before committing accepted outputs. The post notes that standard autoregressive decoding needs 4 sequential decode steps to produce 4 output tokens, while speculative decoding can reduce target-model verification rounds when candidates are accepted. It changes generation by committing multiple tokens per target verification step and discarding later candidates after the first rejection, with observed output-token throughput varying by drafting method, proposal length, and AMD MI300X/MI355X workloads.
OpenAI’s GPT-6 Astra became the first AI model to score a perfect 450 on South Korea’s 2026 CSAT evaluation results posted on GitHub. Astra used 357,000 tokens, the lowest reported total among publicly available models. The report changes expectations for AI efficiency and performance on standardized questions, while experts still warn scores don’t fully reflect real-world problem-solving.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
