X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products, and Applications¶
Edge0 releases 145B and 120B on-device model weights: vendor says peak memory is about 29GB, all benchmarks are self-run¶
Edge0 released Edge0-145B and Edge0-120B on November 12, opening up the weights and the inference framework. The company blog describes them as on-device foundation models for personal computers: 145B total parameters with about 10.12B activated per token, and 120B total with about 8.64B activated, both 48-layer sparse MoE with 512 experts and a 256K context. The GitHub documentation states that both models are derived from Qwen3.8-Flash-Next. Founder Samuel Zeng posted an introduction near the start of the window, saying it can run offline on a laptop and citing figures such as 86.5% on Terminal-Bench 2.1 versus Claude Opus 4.6's 78.2%, about 29GB of memory, and 75+ tokens/s decoding.
This is worth watching because the discussion within the window about "running hundred-billion-plus-parameter models locally on consumer machines" is concentrated here rather than on yet another cloud API. It also spread noticeably further than the official account itself: Samuel Zeng's introductory post has about 601K views, 1,434 likes, 196 reposts, 158 replies, and 1,356 bookmarks; the Edge0 official account's retelling drew far less engagement, about 1,941 views and 38 likes. There are also some hands-on posts, for example marcusyul with about 38K views and 116 likes, and Lacoste0x_ with about 18K views and 120 likes. Engagement data was collected around 08:20 on 2026-11-14 (Shanghai time).
Verified: the Edge0 official blog confirms the November 12 release, the 145B/120B tiers, peak memory of 29GB, prefill of 1100+ tokens/s, decoding of 80+ tokens/s, and other self-run results, as well as scores such as 86.52 on Terminal-Bench 2.1, 90.4 on GPQA-Diamond, and 64.6 on DeepSWE 1.1, and states that the comparisons between Edge0, Qwen, DeepSeek, and Claude come from the company's internal evaluations; the Hugging Face model page provides Q3_K and NVFP4 weights; the model documentation confirms the derivation from Qwen3.8-Flash-Next and the 10.12B/8.64B activated parameters; the license text is the Edge0 Community Model License 1.0 dated November 12, 2026, under which personal, research, internal use, and redistribution are free within the terms, but a "Commercial Model Offering" requires separate authorization and the upstream Qwen license must also be observed.
Assessment: the weights have indeed been released, but "open source" should not be read as Apache 2.0; the leaderboard numbers are currently vendor self-run only, and the blog itself defers reproduction details to an upcoming technical report. Reading this as "already stably matching the strongest cloud models on a consumer laptop" still lacks independent reproduction.
llama.cpp and ggml release 1.0.0: the public API now follows semantic versioning¶
On November 13, llama.cpp and the underlying tensor library ggml released 1.0.0 in sync. This is the project's first major version in roughly three and a half years, and the core change is not a sudden new backend but the removal of already-deprecated public APIs and the announcement that future API/ABI incompatibilities will only come with major version bumps. For developers depending on Ollama, LM Studio, and a large number of local applications, this means the interface boundary is now formally constrained. Author Georgi Gerganov's release post has about 193K views, 3,082 likes, 552 reposts, 145 replies, and 640 bookmarks; Hugging Face's Pedro Cuenca also posted congratulations, with about 7,241 views and 272 likes. Data was collected around 08:20 on 2026-11-14 (Shanghai time).
Verified: the Hugging Face official blog confirms the November 13 release, that llama.cpp and ggml entered 1.0.0 simultaneously, and that breaking changes to the public C API will bump the major version from now on; the llama.cpp v1.0.0 release notes list the removal of deprecated interfaces, promotion of the NextN API into llama.h, the addition of FP8 KV cache and the /v1/audio/speech endpoint, and support for models such as Kimi-K3.1-Pro, Ripple, and xing4_0; the ggml v1.0.0 release notes confirm the parallel cleanup of old APIs and the bump of the RPC protocol to v9.
Assessment: the significance of 1.0.0 lies mainly in an engineering commitment rather than a capability leap. The local inference ecosystem finally has an interface that can be depended on by major version, but downstream applications will still need to adapt to this breaking cleanup.
Bland releases Tofu-1: speech-to-speech for noisy, high-stakes calls, currently only an enterprise research preview¶
Voice company Bland launched Tofu-1 on November 13, positioned as a full-duplex speech-to-speech model, emphasizing listening while speaking and reasoning while speaking under conditions such as vehicles, hospital wards, poor signal, and several people talking at once, and the ability to separate the caller from background sound. Both the company and the founder state that the caller, agent, and background audio in the launch video are all generated by Tofu-1 and are not real customer calls. Founder Isaiah Granet's introductory post has about 308K views, 743 likes, 143 reposts, 58 replies, and 479 bookmarks; the Bland official account has about 75K views and 140 likes. Data was collected around 08:20 on 2026-11-14 (Shanghai time).
Verified: the Bland product page confirms that Tofu-1 is the first version in the company's speech-to-speech series, available to a limited number of enterprise research preview customers, aimed at high-stakes call scenarios such as banking, healthcare, insurance, and government; the page also states that the demo is a model-generated call scenario and does not give latency, word error rate, or comparison benchmarks.
Assessment: this is a clear product-direction announcement, but what can be verified now is only the official capability description and one synthetic demo. It is not yet a publicly callable phone agent, and independent evaluations have not appeared.
Deepnote releases the Agent Initiative Benchmark: testing whether agents speak up when they should not¶
Deepnote released the Agent Initiative Benchmark on November 13, placing 12 frontier models into the Slack of a fictional company called PageMantle and examining, across 50 workplace situations, whether they step in unprompted and whether they stay quiet when they should. The official result is GPT-6 Astra first with 8.97, followed by GPT-6.1 Sol at 8.65 and Claude Fable 5.2 at 8.17; the confidence intervals for Astra and Sol overlap, so this 50-question set cannot separate the two. The cheaper GPT-6 Luna scores 6.38 at about $0.007 per task, while Claude Fable 5.1 scores 7.50 at $3.02 per task. Founder Jakub Jurovych's announcement post has about 299K views, 506 likes, 97 reposts, 85 replies, and 598 bookmarks; the company account's retelling drew very low engagement. A discussion sample, Santiago, has about 35K views and 114 likes. Data was collected around 08:20 on 2026-11-14 (Shanghai time).
Verified: the Deepnote official benchmark page confirms the 12 models, 50 questions, 3-minute time limit, averaging of two LLM judges, as well as the scores in the table above, the cost per task, and the summary that "they speak up about nine times out of ten when they should help, and hold back about six times out of ten when they should stay silent"; the page also states that the next version will expand the number of questions. The code repository is on GitHub, under the Apache 2.0 license.
Assessment: the value of this benchmark lies in the question design — proactive intervention and timely silence — rather than another run up a question-answering leaderboard. But 50 questions is a small scale, the scores come from LLM judges, and the evaluator itself sells an agent workspace, so the numbers should be read as officially self-reported results.
Dwarkesh Patel publishes a long essay revising his timeline: median expectation of an AI researcher on par with top humans in 2028¶
Podcast author Dwarkesh Patel published a long essay on November 13, "The avalanche that hit math is coming for AI development," shifting his judgment on superintelligence from "at least nearly another decade" to "2030, possibly earlier." The median estimate in the essay is that, if labs do not deliberately slow down, an automated researcher comparable to the very best human AI researchers will appear in 2028; he also gives a 25% probability that this step comes after 2032. His announcement post has about 197K views, 2,074 likes, 152 reposts, 68 replies, and 1,373 bookmarks. Data was collected around 08:20 on 2026-11-14 (Shanghai time).
Verified: the original essay confirms the above timeline judgment, along with his citation of METR's finding, after embedding at Anthropic, of roughly a doubling in progress speed, and the published figures from OpenAI and Anthropic on internal AI R&D usage rates. The METR acceleration point was already covered in the November 11 daily, and here it serves only as evidence for his updated judgment, not as a new observation.
Assessment: this is an update to an individual's research judgment, not a new capability evaluation. It sits adjacent to the AI Futures Project's collective forecast on November 11 of "a superhuman AI researcher in 2027," but the framing differs, and it should not be read as an observed capability leap.
OpenAI confirms to Fortune: the standalone robotics team is folded into core research and Applied, new release, not yet high engagement¶
Fortune reported on November 13 that OpenAI has disbanded its standalone robotics team and redistributed the staff to other parts of the company. The company confirmed the change when asked by the reporter, with a spokesperson saying robotics remains a long-term direction and that the reorganization is meant to speed things up rather than to abandon the effort. The report says robotics research was folded into the main research organization reporting to Chief Research Officer Mark Chen, while applied robotics and hardware were folded into the Applied division reporting to President Greg Brockman. Related posts on X drew little engagement, for example TekonBurns's retelling with about 1,017 views and 15 likes, and AIStockSavvy with about 4,887 views and 15 likes. Data was collected around 08:20 on 2026-11-14 (Shanghai time).
Verified: the Fortune report confirms the company's response, the departmental assignments, the first shrinking of the robotics team in 2020, its rebuilding in 2024, the end of the partnership with Figure in February 2025, and on-the-ground details such as OpenAI being registered at CoRL but having an empty booth at the time of the interview. OpenAI did not publish a separate official blog post on this.
Assessment: this is an organizational adjustment, not an exit from robotics. The public information is enough to show that the company prefers to control robots directly with general models, but the report offers no new hard evidence on the number of people reorganized, internal timelines, or whether the 2027 humanoid robot demo can be delivered.
2. AI Compute and Industry¶
Follow-up: former Anthropic pretraining researcher posts an IPO roadshow slide, says the pretraining efficiency curve was a major reason he left¶
Former Anthropic pretraining researcher Jacob Coxon (@hilbertspaess) posted a slide from an Anthropic IPO roadshow on the night of November 13, writing that this chart was a major reason for his decision to leave and expressing surprise that the company made it public. In the post he says the chart shows how pretraining compute efficiency changes over time, with a logarithmic vertical axis, and writes that they saw a sustained "superexponential" trend. The text is cut off after this sentence, with the remainder in the attached image, and a more complete text could not be retrieved from the API this time. The original post has about 260K views, 2,294 likes, 187 reposts, 140 replies, 1,081 bookmarks, and 61 quotes. Data was collected around 08:20 on 2026-11-14 (Shanghai time).
Verified: Coxon's identity can be confirmed from the display name on his X account; he announced leaving Anthropic in September 2026, per the TechCrunch report at the time, so this is not a resignation that happened that day but his later commentary on already-public roadshow material. The launch of the roadshow itself was already covered in the November 11 and November 13 dailies.
Assessment: "superexponential" is Coxon's reading of this log-scale chart, not an official statement attributed directly to Anthropic in the post. The roadshow material had already been made public earlier; the incremental value of this item lies in the researcher's interpretation and its spread, not in the company disclosing a new set of independently checkable efficiency numbers.
3. Chinese AI Ecosystem¶
No Chinese AI news worth a separate item was found in this period. The November 13 daily already covered Doubao Work's (豆包工作) integration of third-party models; there were still retellings within the window, but they did not amount to a new development.