AI Engineering Signal #44
OpenAI's GPT-next disproves 80-year-old Erdős conjecture for under $1,000
Signals
OpenAI's GPT-next disproves 80-year-old Erdős conjecture for under $1,000
AI-generated formal math proofs now require verification infrastructure before trusting in production pipelines.
Latent Space
Gemini 3.5 Flash tops APEX-Agents-AA benchmark
agent routing and model-selection logic needs re-benchmarking against this smaller model now.
Latent Space
CODA fuses transformer blocks as GEMM-epilogue programs
fused kernel approach cuts memory round-trips; test against vLLM baselines before next capacity review.
ArXiv
Meta serves legal notice to Heretic project
audit any local inference stack built on unofficial Llama 4 forks for takedown exposure.
FTC fines Cox Media Group over deceptive AI "active listening" claims
review vendor contracts for behavioral targeting claims before FTC scrutiny lands on your stack.
Simon Willison
Qwen3.6 35B hits 110 tok/s on 12 GB VRAM via ik_llama.cpp
update on-device deployment cost models; consumer hardware local inference is now viable at this scale.
The Take
Capability is outpacing the verification and routing infrastructure most teams have in place — AI-generated formal proofs need audit pipelines, sub-flagship models are beating larger ones on agent benchmarks, and the FTC and Meta legal actions confirm the compliance surface is expanding as fast as the capability surface.
Subscribe
Related Signals