// model scaling

All signals tagged with this topic

Sep 10

When AI Labs Race to Solve Math's Hardest Problems

The competitive pressure among AI labs to demonstrate capability on the Millennium Prize Problems—unsolved mathematical challenges worth $1 million each—marks a shift from theoretical research benchmarks to concrete, externally-validated achievements. Labs are now staking claims on problems with decades of academic pedigree, making capability gains harder to dispute. The rumor itself (whether true or not) signals that labs understand solving canonical hard problems carries political and epistemic weight. It's not just capability—it's credibility. In a field where "whose AI is smarter" increasingly determines funding, talent, and regulatory influence, canonical solutions become a form of proof.

theme-ai cost of compute ai economics model scaling

Jul 14

AI's escalating costs force executives to recalculate the business case

Source: The Register: Biting the hand that feeds

The economics of large language models—particularly the inference costs of running tokens at scale—are creating genuine friction in boardrooms where the ROI math no longer works. CFOs are discovering that the computational cost per transaction makes many proposed AI applications uncompetitive against traditional software. The industry will likely segment sharply between a small number of high-volume, low-margin players (cloud giants, search) who can absorb token costs and everyone else scrambling for narrow, defensible use cases where AI's margin contribution justifies the infrastructure spend.

theme-ai llm training model scaling hardware constraints

May 25

Huawei's Tau Scaling Law Bypasses Transistor Miniaturization Race

Source: The Next Web

Rather than compete on transistor density—where US sanctions have blocked access to advanced nodes—Huawei is optimizing for signal propagation delay, a physics constraint that can be engineered through architecture and packaging instead of fab precision. This bypasses the need for cutting-edge manufacturing. China can build competitive AI chips using older, domestically available nodes if the architecture is efficient enough. Geopolitical constraints are forcing genuine technical innovation in chip design, not just nationalist substitution.