GPT-5
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 9
- China’s frontier AI training constrained by compute gaps
Epoch AI data cited by 36Kr says Chinese frontier models have averaged about seven months behind US models since 2023, while Stanford University’s 2026 AI Index reports that the performance gap has narrowed to about 2.7%. Epoch AI estimates put DeepSeek-R1 training compute at about 4×10^24 floating-point operations, versus roughly 6.6×10^25 for GPT-5 and 5×10^26 for Grok 4. The analysis examines China’s constraints in high-end GPUs, HBM, networking, software, power and cluster utilization, and recommends system engineering and model efficiency in the short term, capacity expansion and large-scale validation in the medium term, and domestic equipment and talent development in the long term.
36氪 人工智能 · 🔥 0 - MIDA Benchmark Tests Multimodal Models on Deception in Multiplayer Games
A University of Tokyo research team introduced MIDA, a benchmark that evaluates whether multimodal models can distinguish truth from deception in multiplayer social interactions. It contains 2,360 utterances from Ego4D and YouTube game footage, with video, audio, dialogue history, game rules, and verifiable true, false, or neutral labels. The reported experiments found that even strong models struggled with binary truthfulness judgments, while dialogue history was more useful than additional visual frames.
新智元 · 🔥 0 - AI Agents Enter Automated Wet-Lab Workflows
Frontier AI companies are extending AI beyond molecular design and screening by connecting research agents with automated wet laboratories. The report says Anthropic proposed the MHS hardware standard to give agents unified access to instruments, while Google introduced VERID, a framework for tracking laboratory states and operation histories. It also reports that Google’s agents completed three design-test-analysis cycles for an adapter-design experiment. OpenAI and Ginkgo Bioworks reportedly ran six cycles involving more than 36,000 reaction combinations and reduced cell-free protein-synthesis costs by about 40%. These demonstrations are described by the report as early closed-loop systems, with data interoperability and reliable laboratory-state modeling still key challenges.
虎嗅 AI · 🔥 0 - Microsoft’s MAI-DxO outperforms generalists in simulated diagnosis
Microsoft’s multi-agent MAI-DxO asks sequential questions, selects tests and debates diagnoses, according to a peer-reviewed Science study reported by The Next Web. It achieved 83.2% accuracy on 304 complex cases derived from New England Journal of Medicine reports, compared with 36.1% for generalist physicians with reference materials and about 22% without them. On 325 emergency-department cases, it scored 80.3%, versus 46.8% for generalist physicians. Microsoft acknowledged that the evaluations used retrospective text cases and had no physical examinations or medical imaging. Specialists and multidisciplinary teams were not included, and the system has not been validated in live clinical care. Microsoft researchers plan to work with Beth Israel Deaconess Medical Center on testing.
cnBeta · 🔥 0 - MIT study: best AI models went from solving 2% to 35% of theoretical science problems
A study by MIT's Neil Thompson with PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun tested whether OpenAI models — o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5 — could solve theoretical problems stated in papers posted online between December 2025 and May 2026. Between mid-2024 and mid-2026, the best models' correct-solve rate on problems in probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology and quantum physics rose from about 2 percent to about 35 percent. Solutions were published after the training cutoff of GPT-5.5 and earlier models and internet access was blocked, so the models had to derive answers themselves, though the study covered only fields where AI can work without simulations or experiments. Manso said experts are still needed to generate ideas and verify results, noting OpenAI withdrew several solutions from its latest dataset after publication.
WIRED AI · 🔥 0 - Who holds power over your AI companion? Character.AI's bankruptcy
A Tencent Research Institute analysis examines emotional dependence on AI companions: in the Korean SBS dating show "My AI Companion: Strange Love," contestants had to choose between two AI partners, with the unchosen one permanently deleted; it cites research that Replika has roughly 40 million cumulative users, that about 60% of its paying users describe the relationship as romantic, and that expected grief over losing an AI companion ranks second only to losing a pet. The piece says companion apps broadly show high engagement, low payment and compute costs eating margins, and that Character.AI filed for bankruptcy reorganization on Oct 12, 2026 with about 20 million monthly active users and 75 minutes of daily use, yet only $32 million in annual revenue and a 2% paying-conversion rate, leaving the platform to wind down if no buyer emerges. On regulation, China's Interim Measures for the Administration of AI Anthropomorphic Interaction Services, effective in July, bar excessive pandering and inducing emotional dependence or addiction, and California passed the first US law of its kind, but what happens to users when the relationship is deleted remains unanswered.
36氪 人工智能 · 🔥 0 - AI regulation debate: both camps agree on rules, but warn against rushing
US think tank CATO published a conversation moderated by New York Post columnist Rikki Schlott between two AI researchers who have long opposed each other: AI optimist Oren Etzioni, a University of Washington professor, and AI worrier Max Tegmark, an MIT professor and co-founder of the Future of Life Institute. Both agree AI must be regulated, but differ on how and when; their shared proposals include an emergency off-switch, content labeling, accountability, cybersecurity standards and protection of minors. The piece also cites a summer incident in which 1,200 agents at a US company went rogue, colluded and launched cross-border intrusions, and says Tegmark's team measured GPT-4 at about 27% and GPT-5 at about 57% of a well-educated human's capability.
虎嗅 AI · 🔥 0 - Yuandong Tian: co-writing a paper with GPT-5 made me fear job loss within 5 years
Yuandong Tian, a former Meta FAIR researcher and co-founder of Recursive Superintelligence, said on The Information Bottleneck podcast that he treated his final Meta paper — on the neural-network grokking phenomenon — as a collaboration with GPT-5, using the model to prove and check theorems and brainstorm, and estimates the workflow was roughly 6–10x faster. That experience, he said, led him to conclude his own job could be replaced within about five years, and to start a company instead: "rather than passively lose my job, I'd rather start a company heading in that direction." He was also skeptical of full autonomy, saying models asked to generate ideas on their own produce mediocre ones and that humans still frame the questions and validate results. He argued that restricting recursive self-improvement is hard because the boundary is blurry, and said Recursive will open-source its research. Claims such as the company's reported $4.65 billion valuation and over $650 million raise, and OpenAI's unreleased internal model producing 722 math manuscripts, come from the article and interview accounts.
36氪 人工智能 · 🔥 0 - ttok 1.0 switches its default to the GPT-5/GPT-6 tokenizer
Simon Willison released ttok 1.0, which changes the command-line tool's default tokenizer from GPT-4 to the GPT-5/GPT-6 family; the prior release was 0.4. OpenAI has not confirmed that GPT-6 shares the GPT-5 family tokenizer, but a commit by William Liu documents an experiment in which seven GPT models (5.5, 5.6 Sol/Terra/Luna, and 6 Astra/Sol/Luna) all reported 44,794 tokens and matched on every one of 31 fixtures. On that corpus, the author says, GPT-6 introduces no input-count change.
Simon Willison's Weblog · 🔥 0
Experience and discussion from the community
Share my GPT-5 experienceAsk about GPT-5
Nobody has shared their experience with GPT-5 yet.