Terence Tao published a full transcript of a worki, koala73/worldmonitor debuted at #1 on GitHub Trend, GigaToken, a Rust/SIMD-optimized BPE tokenizer, cl

Terence Tao published a full transcript of a working session with a 'Pro' tier GPT-5.6 model, probing an alternate construction related to the Jacobian Conjecture counterexample announced July 21 by Levent Alpoge with Anthropic's Fable 5. Tao didn't discover anything new and neither did the model -- it verified polynomial identities and Jacobian computations Tao directed it toward -- but publishing the raw prompt-by-prompt session as the primary artifact, rather than a polished paper, is a new instrument for judging AI mathematical assistance at the highest level of expertise.

koala73/worldmonitor debuted at #1 on GitHub Trending with +4,139 stars/day (69,556 total) -- more than 5x the next entry and the largest single-day velocity this research stream has logged this month. It's a free, self-hostable real-time global intelligence dashboard fusing 56+ data layers (conflict events, ship/flight tracking, pipelines, markets, wildfire detection) onto a 3D globe, reportedly used by 2M+ people across 190+ countries.

GigaToken, a Rust/SIMD-optimized BPE tokenizer, claims up to 1000x faster throughput than HuggingFace's tokenizer and OpenAI's tiktoken (24+ GB/s on a 144-core EPYC, 8+ GB/s on an Apple M4 Max), by replacing regex-based pretokenization with vectorized token-boundary detection and aggressive caching. The gains are concentrated in batch/pretraining workloads and are author-benchmarked only so far.

Cactus Hybrid adds a lightweight, calibrated confidence head to an on-device Gemma 4, cutting overconfident wrong answers by ~34% on TriviaQA (confidence averages 0.82 on correct answers, 0.29 on incorrect ones) at a 3-5ms/token cost -- enabling hybrid routing where only 15-35% of queries need escalation to a larger cloud model to match its quality.

OpenAI raised its planned AI infrastructure spending to $750B through 2030, up 25% from an earlier $600B estimate, anchored by Project Camellia -- a $20B, 1,400-acre data-center campus near Savannah, Georgia contracting 3.2 GW of power from Georgia Power -- marking a shift from leasing cloud capacity toward OpenAI directly designing and operating its own facilities.

tirth8205/code-review-graph posted +882 stars/day (25,459 total) today -- its first slowdown after five consecutive days of GitHub Trending acceleration that peaked at +1,925/day (24,764 total) on July 22.

ayghri/i-have-adhd, the agent-response-structure skill that debuted July 22, held #3 on GitHub Trending for a second consecutive day (+1,699 stars/day, 8,520 total) -- closer to sustained growth than a single launch-day spike.

Media3/ExoPlayer's OTT stack saw only routine maintenance since July 16 (audio-session race-condition fixes, AV1-Dolby-Vision codec artifact fixes via the dav1d decoder, HLS playlist propagation bugs) -- no new consolidation, DRM, or gateway-control development to report this Thursday's streaming-tech rotation.