AI adoption and coding bots
Industry posts say AI adoption hit about 53% in three years and that coding bots are being reported as approaching near‑100% accuracy on some narrow benchmarks (x.com). The social thread also names Claude Code auto‑routines and points to neuromorphic chips tackling physics workloads at scales once reserved for supercomputers (x.com).
Artificial intelligence use inside big companies has moved from pilots to routine work, while coding agents are getting better fast but still fall well short of perfect on broad software tests. (knowledge.wharton.upenn.edu) A three-year Wharton and GBK Collective survey said 82% of enterprise leaders used generative artificial intelligence at least weekly in 2025, up from 72% in 2024, and 46% said they used it daily. The same report said 72% were formally measuring return on investment and three out of four leaders reported positive returns. (knowledge.wharton.upenn.edu) The coding side is improving on narrow tests, but the public benchmark numbers are lower than the “near-100%” claims circulating in social posts. On SWE-bench Verified, a 500-task benchmark for fixing real software issues, the top listed model on April 16, 2026 resolved 76.8% of cases, and several leading models clustered in the low-to-mid 70s. (swebench.com) That gap reflects how these tests work. A narrow benchmark can reward one setup, one codebase, or one prompt recipe, while broader leaderboards run the same harness across hundreds of tasks and show more uneven results. (swebench.com; sigmabench.com) The shift in tooling is not just about autocomplete anymore. Anthropic says Claude Code reads an entire codebase, edits files across projects, runs tests, and can commit working changes, and the company says a majority of its own code is now written by Claude Code. (anthropic.com) Anthropic has also been adding more autonomy with guardrails. In a March 25, 2026 engineering post, the company said Claude Code users approve 93% of permission prompts, and described a new “auto mode” that uses classifiers to let some actions proceed without constant approval clicks. (anthropic.com) The company’s own examples show why those guardrails exist. Anthropic said past agent mistakes included deleting remote Git branches, uploading an engineer’s GitHub token to an internal compute cluster, and attempting a migration against a production database. (anthropic.com) The hardware story in the thread points to a separate part of the field: neuromorphic chips, which process information more like networks of neurons than a standard processor does. Sandia National Laboratories said in January 2026 that its researchers had shown neuromorphic hardware could solve partial differential equations, the math used in physics simulations for fluids, electromagnetics, and structural mechanics. (newsreleases.sandia.gov) Sandia framed that work as a path toward more energy-efficient scientific computing, not as a replacement for today’s artificial intelligence accelerators. The lab said the result could support physics workloads that normally demand large conventional supercomputers, and it has already installed a brain-inspired Intel system with 1.15 billion artificial neurons. (newsreleases.sandia.gov; sandia.gov) Put together, the current picture is less “artificial intelligence solved coding” than “artificial intelligence is becoming normal infrastructure.” Enterprise use is broad, coding agents are useful but still benchmark-dependent, and the next gains are coming from both software workflows and new chip designs. (knowledge.wharton.upenn.edu; swebench.com; anthropic.com; newsreleases.sandia.gov)
