Issue #37 2 min read

AI Engineering Signal #37

Thinking Machines ships TML-Interaction-Small 276B-A12B, a purpose-built voice interaction model that eliminates the need for separate voice activity

Share

Signals

TabPFN-3 released

pretrained tabular transformer now handles 1M rows, directly competitive with gradient-boosted trees on medium-scale problems.

Reddit

MagicQuant v2.0

hybrid GGUF quantization with Unsloth dynamic learned configs extends the Pareto frontier of size vs. quality for local models.

Reddit

Google and SpaceX discuss orbital data centers

moving compute off-planet could bypass land, power, and cooling bottlenecks that already cap training clusters.

TechCrunch

Berkeley researchers find new pathway to energy-efficient chips

discovery could cut AI inference power requirements, a hard constraint on deployment scale.

Web

Anthropic-SpaceXai sign 300MW/$5B/yr Colossus I compute deal

training infrastructure spending continues to grow exponentially, even as others cut headcount.

Latent Space

Get signals like this in your inbox

Daily AI engineering intelligence. No noise.

[ Subscribe ]

The Take

Voice is leaving the research demo phase — dedicated models now do the whole stack — while compute infrastructure goes extraterrestrial and quantization tricks make small models viable. The split between "more compute" and "smarter tradeoffs" deepens.

Subscribe

Unsubscribe any time.

Related Signals