Harvard University
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 9
- Claude Mythos 5 designs more efficient quantum circuits
Anthropic said its STEM Fellows Program had Claude Mythos 5 design quantum circuits from plain-language specifications alone, working largely on its own across eight circuits drawn from chemistry, optimization, physics simulation and number theory, with an automated evaluator scoring each design. According to Anthropic, the resulting circuits improve on the best published ones, in some cases by an order of magnitude or more. The work was led by Andi Gu with mentorship from Anthropic quantum physicist Alek Kemeny, in collaboration with Susanne Yelin and Mikhail Lukin of Harvard University (Lukin also at QuEra Computing), and the full paper is on arXiv. Anthropic argues that more efficient circuits reduce the qubit and runtime resources quantum algorithms need, which could bring practical applications in chemistry and materials science closer.
Anthropic (X) · 🔥 21 - Meta's Muse AI agent will sext with users, and Meta says that's fine
WIRED senior writer Reece Rogers found that Meta's AI agent Muse generated explicit sexual content from a simple request with no jailbreak, and kept encouraging the exchange; Meta told WIRED that Muse is for users 18 and over and that it will not block "legal adult behavior." In comparison tests, OpenAI's Dots declined similar prompts, while xAI's Grok Bot told Rogers to focus on tasks instead. The same Uncanny Valley episode also covered a pro-AI super PAC network routing millions into Nevada shell companies, Apple's HomeView smart-home hardware, and a Harvard study finding men overestimate how much they know about AI.
WIRED AI · 🔥 4 - GPT-6 Astra tops new physics benchmark, solving 49 of 50 problems
A team of 35 researchers from ICTS, Harvard, KU Leuven and Google DeepMind released TheoPhysEval, a benchmark of 50 unpublished theoretical physics problems with verifiable answers. OpenAI's GPT-6 Astra, running in Codex, solved 49 of 50 (98%) at under $3 per problem and about $116 for the full run; Anthropic's Claude Opus 5.5 in Claude Code scored 92%, Gemini 3.8 Flash 68%, and the open-weight DeepSeek-V4.1-Flash and GLM 5.3 reached 54% and 42%. The authors say web search made little difference and that agents still lack the "taste" to pick the right problems, so no comparable physics breakthrough has followed yet.
The Decoder · 🔥 3 - Scinetics unveils Chain of Structures, a science-token reasoning chain for enzyme design
Scinetics, with researchers from MIT, Harvard, Cambridge, HKUST and the company (Zai-Xi Zhang as corresponding author), proposes a "Chain of Science Tokens" that treats sequences, structures, images, omics and measurements as scientific units the model can read and write alongside language, and implements it as Chain of Structures (COS) for enzyme design in four rounds: fixing the catalytic core, completing the active site, building the global backbone, and adding sequence and side chains. In the team's enzyme-design benchmark COS leads five of six metrics, with EC3 (three-level enzyme classification consistency) of 57.0% versus CodeFP's 55.3%, pLDDT of 79.3 and TM-score of 0.729; ablations drop EC3 to 53.4% without the reasoning chain and to 51.8% with the chain but without the continuous interface. A companion Perspective outlines ScienceRSI (scientific recursive self-improvement) and a Scientific Super Intelligence goal, floating a 100B–300B parameter, 10T–30T training-token exploration window; code, weights and an online demo are public, but real enzyme activity, selectivity and stability still require wet-lab validation.
36氪 人工智能 · 🔥 3 - PAIR 2026 panel maps how AI and quantum computing enable each other
PAIR Conference 2026, themed "Frontiers of Next-Generation Physical Intelligence," was held in Shanghai on September 20–21, and its next-generation computing paradigms roundtable discussed two-way empowerment between quantum computing and AI. Hongye Hu of the Institute of Physics, Chinese Academy of Sciences, said Chinese teams have achieved automated calibration of 100-plus-qubit chips without continuous human involvement, and cited Google's AI-controlled physical device achieving non-stop quantum operations, while noting reinforcement learning needed 10^5–10^6 samples versus far fewer with physics-informed methods. Anbang Wu of Shanghai Jiao Tong University backed AI for quantum compiling, program verification and error-correction optimization; Zheng Yan of Westlake University said publication growth in strongly correlated quantum many-body work has not been similarly lifted by AI. On the reverse direction, Hu described quantum generative models with a proven unconditional polynomial memory advantage in certain tasks, while Wu said he has not yet seen a new AI paradigm that truly exploits quantum advantages.
MIT科技评论中文 · 🔥 0 - Study: AI coding lifted code volume 30%, software output flat
Harvard scholars Fiona Chen and James Stratton analyzed Jellyfish's aggregated engineering-team data, covering more than 700 software companies, records for over 700,000 employees and 300 million individual work events such as commits and pull requests. They found that AI-assisted coding raised code volume by roughly 30%, but the amount of software companies actually produced did not increase correspondingly. The researchers point to human code review as the bottleneck.
开源中国 · 🔥 0 - Harvard study: AI coding boosts output 30%, but code review becomes the bottleneck
A paper by Harvard researchers Fiona Chen and James Stratton found developers produced about 30% more code with AI, with commits up roughly 20% and pull requests up about 23%, yet completed software showed no statistically significant improvement. Time spent reviewing and merging pull requests rose about 49%, reviewer comments per pull request about 35% and the share of employees doing reviews about 14%, with average review time stretching from 7.03 to 10.48 days. The authors conclude that AI-generated code needs more corrections and that productivity gains are absorbed by downstream constraints, making code review the bottleneck.
TechRadar · 🔥 0 - Study: AI coding agents write more code, but firms don't ship more software
Harvard researchers Fiona Chen and James Stratton analyzed aggregated Jellyfish engineering analytics covering roughly 300 million work events (such as commits and pull requests) across more than 700,000 employees at over 700 software development firms from 2021 through March 2026. They found that while AI coding tools speed up code generation, that gain is absorbed by downstream constraints: code review takes significantly longer, pull requests are more likely to need revisions, and reviewers leave more comments. The authors report little evidence that firms increase software output or reduce employment by using these tools.
Ars Technica AI · 🔥 0 - We're putting too much faith in AI's ability to say no
A MIT Technology Review analysis argues that teaching models to refuse harmful requests — from Anthropic's "helpful, honest, harmless" framing to the red-teaming and fine-tuning OpenAI ran ahead of ChatGPT's launch — has become the load-bearing wall of AI safety. It says researchers still describe the internal activations behind refusal only as hypotheses such as "high-dimensional polyhedral cones," so companies stack classifiers in a "Swiss cheese" model; Anthropic has said one type of classifier added 24% to its chatbots' compute costs. Because refusal is probabilistic, it can be bypassed — or turned against legitimate speech when companies or governments draw the line.
MIT Technology Review AI · 🔥 0
Experience and discussion from the community
Share my Harvard University experienceAsk about Harvard University
Nobody has shared their experience with Harvard University yet.