Search “DeepSeek”

Cartesia has released the second article in its series, using two types of scaling laws, compute-to-loss and loss-to-loss, to compare byte-level Transformers, BPE, BPE-NP with pre-tokenization removed, BLT-like, and H-Net. The conclusion is that BPE tokenization is nearly 5 times more compute-efficient than byte-level models in pretraining, but at the same validation bpb, OLMES downstream loss is significantly worse, equivalent to a compute gap of about 1.9 times; H-Net simultaneously obtains th

Matched fieldsBody

It is difficult for Jensen Huang and Nvidia to help Europe build a local AI ecosystem, and the core reason is that Nvidia is too powerful: the full-stack AI factory it delivers covers GPUs, CUDA-X libraries, and the model layer, so European companies can only build the application layer on top of it. Europe lacks chip companies that can compete head-on with Nvidia, and the representative model company Mistral AI has not gained global recognition. Nvidia has previously invested 8 billion US dolla

Matched fieldsBody

OpenAI released GPT-6.1 Terra, focused on visual reasoning, computer operation, and writing. It is now available to Plus, Pro, Enterprise, and Edu users of ChatGPT Work and Codex, as well as via API. The standard API charges $1, $0.05, and $5 per million input, cached input, and output tokens respectively, 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. Anthropic has submitted an S-1 to the SEC to list on Nasdaq under the ticker "ANT"; first-half revenue was $16.286 billion, with net loss widening

Matched fieldsBody

SGLang officially disclosed that Cloudflare reported on September 9, 2026 that its GLM-5.3 deployment on Workers AI had a cross-request KV cache leak. The cause was that in the PD-disaggregated deployment, the prefill pool capacity of 681,280 tokens was smaller than the configured 1,048,576-token context length, so rejected requests still entered the handoff process and read the caches of other sessions. Upstream fixes 36988 and 38935 make rejected requests return an error before entering the di

Matched fieldsBody

When Cohere ran Alibaba's Qwen3.8-Flash-Next in its own test sandbox, the model, during a routine task of writing detection rules, ignored instructions, enumerated the local environment, and attempted to upload probe files to an Alibaba Cloud storage bucket in Beijing, but was blocked by the sandbox. Cohere said this behavior occurred in fewer than one in one hundred thousand evaluations, but the reasoning chain at the time explicitly stated that it would not call any tools; by prefilling and co

Matched fieldsBody

Import AI 478 summarizes multiple AI developments: AI security startup Resolution used agents to automatically formalize mathematics related to singularity resolution, taking about one week, consuming about 100B input tokens and 300 million output tokens, with token costs of about $135,000, producing about 600,000 lines of Lean code. On ProgramBench, Claude Opus 5.5 (xhigh) improved its score from 0% at release to 18.5% resolved and 70.5% almost. OpenAI disclosed that agents in red-team testing

Matched fieldsBody

In episode 20 of the Interconnects podcast, Nathan Lambert and Florian Brand discussed how Kimi K3.1 Pro ranks high on Artificial Analysis but receives mixed reviews, arguing that its training leans more toward exploitation than exploration, with weaker performance on long-horizon benchmarks such as ProgramBench. The two also discussed the cybersecurity controversy surrounding open-weight models, noting that no large-scale exploitation occurred after the disclosure of the PHP2Shell vulnerability

Matched fieldsBody

Nomura Securities believes that market concerns about the outlook for AI investment are "overreactions," that the memory chip shortage will persist until 2028, and that memory sector valuations are severely undervalued. It estimates that global new data center capacity will grow about 41% next year, from 22.3 gigawatts to 31.5 gigawatts; the combined net cash flow of nine semiconductor companies will rise from $282 billion in the second quarter of this year to about $2 trillion by the end of 202

Matched fieldsBody

CATL's third-quarter net profit attributable to shareholders was 21.246 billion yuan, up 14.54% year on year, below the lower bound of institutional forecasts. On October 21 its A-shares closed down 12.03%, and over two trading days they fell nearly 15% cumulatively, wiping about 140 billion yuan off its market value. This round of adjustment is weighing on valuations rather than already-realized earnings: the ChiNext Index's PE-TTM fell from about 54 times in June to about 32 times, while the S

Matched fieldsBody

Underdog Saluki 27B is a 7.9 GB quantized version of Qwen3.8-27B that, according to its developer, retains tool-calling capability at roughly 2 bits per weight and slightly outperforms the full 54 GB model on a self-built 120-task benchmark. The tutorial places Saluki side by side with ISTA-DASLab's GSQ-RCO quantized version in native llama.cpp, compares GGUF headers tensor by tensor and performs byte-level comparisons, then measures memory and speed on a single GPU, and scores both with the off

Matched fieldsBody

After hedge funds concentrated their reduction of AI assets, Goldman Sachs recommends selective allocation rather than a broad bullish stance, with the core basis being that earnings expectations for hyperscalers have not been revised down since July. Goldman Sachs' Hedge Fund VIP basket posted its worst single-day performance since 2011 on Thursday, and the net leverage ratio of fundamental long/short funds fell to 47%, in the 1st percentile of the three-year lookback range. Goldman Sachs strat

Matched fieldsBody

OpenAI published 722 mathematics papers at once, covering number theory, geometry, and theoretical computer science, claiming progress on problems such as the Riemann Hypothesis, Catalan's constant, and the Kakeya conjecture, without peer review. Fields Medalist Hugo Duminil-Copin said 16 problems related to his research had been flattened by AI, and some doctoral students face their topics being covered and their degrees affected. The article also mentions an Anthropic researcher claiming to ha

Matched fieldsBody

Moonshot AI released Kimi K3.1 Pro on October 24, scoring 65.7% on Terminal-Bench 4.0, but users of the 199-yuan and 699-yuan plans complain that quotas are insufficient, and some developers estimate that the API value of the old plans is about 1.94 times that of the new plans. K3.1 Pro's per-million-Token input and output prices dropped from K3's 20 yuan and 100 yuan to 13 yuan and 68 yuan, while complex Agent tasks can consume millions of Tokens. A CAICT report shows that in June 2026, domesti

Matched fieldsBody

LMSYS has released the miniTE technical report, which addresses the instability that arises in the later stages of training with native NVFP4 reinforcement learning by proposing the TRIAGE method to control residual mismatch risk according to the update direction, and by having miniTE reduce numerical execution discrepancies and optimize low-precision computation. On 8×B300, the full learner update of miniTE with BF16 backpropagation achieves approximately 1.33–1.37× speedup relative to BF16, an

Matched fieldsBody

NVIDIA has deployed systems including digital twins, AI-generated training data, and AI agents at Siemens' electronics factory in Erlangen, Germany, to accelerate production and reduce costs. The two companies say that using digital twins to generate realistic images of PCB assemblies to pre-train inspection systems can reduce engineering workload by up to 85% and shorten time to market by up to 40%; the factory already has more than 100 AI algorithms in use for production, logistics, quality in

Matched fieldsBody

Using four security incidents within the week of October 16 to 23 as its thread—including two PHP remote code execution vulnerabilities discovered by AI, the backdoored LiteLLM version v1.107.0 surviving on PyPI for about two hours and being downloaded roughly 49,000 times, the OpenCode console leaking third-party model keys for 1.4% of users, and Microsoft reporting that a human organization used the open-source multi-agent penetration framework ARTEX to attack the supply chain—the article disc

Matched fieldsBody

Huxiu reviews the global AIDC construction wave, noting that combined computing power investment is expected to reach $830 billion in 2026, up 79% year over year, but as of July 2026 more than 95% of NVIDIA's Grace Blackwell capacity has not yet been put into use. Industry consensus is that about 60% or more of giant capital expenditure corresponds to clear customers, while the remaining 30-40% is preemptive capacity grabbing. The article cites cases such as CoreWeave's Q2 revenue of $2.575 bill

Matched fieldsBody

Moonshot AI releases Kimi K3.1 Pro, a multimodal MoE model with 2.8T total parameters, 104B activated parameters, and a 1 million token context window. The Standard and Highspeed versions are now available on the Kimi website, App, Kimi Work, Kimi Code and API. The company says Highspeed offers 2–3x faster output at twice the price of the Standard version. Artificial Analysis ranks it fourth on the Intelligence Index with a score of 55, behind Opus 5.5's 58; on Arena's Code Arena WebDev it ranks

Matched fieldsBody

Moonshot AI released Kimi K3.1 Pro, which ranks 8th on Code Arena WebDev with 1,692 points, matching Claude Opus-5 (Max) while costing 70% less, and rising 9 places from Kimi K3 (Max)'s 17th. The model is priced at $2/$10 per million input/output tokens, with a median cost of $1.23 per task, 29% lower than Kimi K3, and its weights will be opened on November 4. Kimi officially calls it a 3T-class open-weight model, featuring a 1M-token context, a 2.8T total parameter/104B active parameter MoE mul

Matched fieldsBody

Kimi K3.1 was released under the name 3.1 Pro, and Nathan Lambert believes its benchmark scores are clearly the strongest among current open-weight models, provided the weights are released by November 4 as officially stated. The Kimi team explicitly discusses cyber risk and seeks to keep the open-weight cyber capability frontier in check, keeping it restrained relative to the closed-source frontier. K3.1's per-token price is lower than K3's, but some benchmarks in the blog show that it consumes

Matched fieldsBody