posts · 2026-08-06

OpenAI dropped a blog post yesterday breaking down ChatGPT usage across its 1B users by country. The headline claims: at work, people are twice as likely to use ChatGPT for "doing" (coding, analysis, creating) vs. "asking"; Latin America, Africa, and Oceania are closing the per-capita adoption gap; multimedia messages hit 7.8% globally, over 10% in Brazil and Colombia; and the 35+ age group is growing in nearly every country. Both sources pull from the same official blog, just with different emphasis. OpenAI's own title pushes the "asking to doing" narrative; aihot highlights the older-user growth. Not a contradiction—more like two angles on the same press material. Where I'd discount: the data covers only individual-tier accounts (Free, Go, Plus, Pro), not enterprise or API usage, so the "work" label is self-reported context, not org-level adoption. The ranking shifts compare just Q1 to Q2 2026—one quarter isn't a trend yet. I haven't seen independent verification of the raw dataset, so treat this as OpenAI's own market signal for now.

The useful bit here is flipping the instinct to dump more context into the model. OpenAI runs Codex offline to read pipeline code and extract real table semantics—keys, grain, join logic—then structures that into descriptions. Online RAG retrieves those pre-curated descriptions instead of raw metadata and query logs. Emma Tang's point: less but more accurate context works better. I'd discount this a bit. No accuracy or ablation numbers are disclosed, and the post doesn't spell out refresh cadence or how often live schema checks actually fire. The model still overconfidently miscalculated ChatGPT active users as 5 million, so clean offline context doesn't fix everything. The takeaway isn't that RAG is dead—it's that the retrieval target got upgraded. If you're running an internal data agent, checking whether table semantics still live only in pipeline code might save more time than tuning chunk size.

This piece caught my eye because it nails a real engineering cost problem: in an alert storm, firing an agent investigation for every single alert means thousands of API calls in minutes. Datadog's Mambark model—roughly 97M parameters—claims to narrow ~10B daily security events down to ~10K candidates before handing off to an LLM agent. That's six orders of magnitude. The numbers are vendor-reported, so I'd discount them a bit, but the architecture is solid. The trick isn't using a small model for generic text classification. It's unsupervised next-event prediction with a Mamba architecture: the log stream is treated as a system state sequence, and the model scores how surprising each event is given what came before. When reality deviates from the expected sequence, it flags the event. This fits better than vector retrieval because nobody is typing queries into a live log stream. The article traces the lineage cleanly—from DeepLog's LSTM in 2017 to Mamba SSMs in 2023—and offers practical steps: use Drain for log template parsing first, then pick a filter tier based on daily event volume. What's missing is independent benchmarks and false-negative rates across industries. But as an architectural principle, putting "decide what's worth investigating" ahead of "make each investigation cheaper" is the right call.

This piece is worth reading because it reframes fine-tuning clearly: it's no longer about injecting knowledge or boosting reasoning, it's cost engineering. FermiSense fine-tuned Qwen3.5-9B for e-commerce review, dropping cost from tens of dollars to $0.50 per 1k calls. Intercom's small customer-support model hit 73.1% resolution at one-fifth the cost of GPT-5.4. On the vision side, a DINOv3 classifier trained on just 839 reviewed samples delivered AP 0.9731 with zero API cost. I'd discount the vendor-reported scores a bit—most lack third-party reproducible test sets. FermiSense's evaluator has self-bias risk, Harvey uses proprietary benchmarks. That's why the article's decision matrix is the useful bit: fine-tuning pays off above ~50k daily requests with automatically verifiable outputs. Below that, Prompt Caching plus RAG is the better bet. The stat that 40% of teams now use hybrid routing—small models handling 80-95% of requests, hard cases routed to top APIs—says more than any single high score.