Lesen

On September 22, Loongson Technology released the first software version of its Loongson Accelerated Computing Platform, targeting the LG200 GPU cores integrated in the 2K3000 and 9A1000 chips, covering drivers, compilers, operator libraries and an inference engine.

The platform supports both OpenCL 3.0 and CUDA programming interfaces, and uses its in-house LacInfer engine as an ONNX Runtime execution backend, so ONNX models exported from PyTorch or TensorFlow can be deployed without code rewrites. Operator libraries include assembly-level optimizations for FP32 and INT8 GEMM workloads.

The company says the software already serves early customers and underpins agent development for embodied devices on the 2K3000 and 9A1000. Note that inference speed and accuracy-loss claims are Loongson's own figures with no third-party benchmarks, and the 9A1000 graphics card is not expected on sale until the first half of next year per the company's earlier statements.

Inspur announced the Yuannao SD200 Ultra supernode AI server at AICC2026 on September 21, built on domestic AI chips. The company says a single node can host the 2.8-trillion-parameter Kimi K3 model and supports frontier models up to 10 trillion parameters.

The system tightly couples 128 domestic AI chips with 8TB of unified-addressable memory and 64TB of system memory. Inspur claims token generation latency below 5.85ms, equivalent to 170 tokens/s per user and five times the industry average, plus a 3.5x reduction in AllReduce communication time.

Inspur also launched the HC2000 compute unit the same day, claiming 10x token throughput per unit of investment. Note that all performance figures are Inspur's own claims; the 'industry average' baseline is unspecified, and real-world results await third-party testing.

XPeng began pushing the new second-generation VLA release, XOS 6.3.0, on September 22, with single-Turing Max vehicles upgraded to the second-generation VLA in the same rollout.

The company says the release introduces the Infini-VLA long-horizon architecture, which it claims remembers the previous 30 seconds of driving context, and puts the X-Foresight predictive world model in a production car for the first time, claiming it can forecast surrounding road users' behavior up to 6 seconds ahead.

All capability descriptions are XPeng's own claims with no independent testing. For delivered Turing Max owners this is an OTA they can receive now; for anyone tracking the driver-assistance race, the numbers that matter will come from user testing and incident data, not launch copy.

Base44's Superagent can now make and receive phone calls on a user's behalf, according to a press release the company shared with TechRadar Pro.

Users give the agent a goal and context; it then contacts businesses or customers, asks questions, collects information, navigates supported phone menus, and reports back. The vendor lists use cases such as rescheduling appointments, requesting vendor quotes, and retrying unanswered calls on a schedule.

This is a vendor announcement with no independent testing. Call success rates, pricing and the exact launch date are undisclosed; whether it handles real phone systems will show in actual usage.

NVIDIA's DSX Ready qualification program went live today, covering battery energy storage systems (BESS) and cooling distribution units (CDU) that sit alongside DSX AI factory racks — two classes of hardware NVIDIA does not make itself.

The first qualified CDU vendors are LG Electronics, LiquidStack and Vertiv; the first qualified BESS vendors are Hitachi Energy, LG Energy Solution and Tesla. CDU vendors may self-qualify against NVIDIA's specifications with final NVIDIA approval, while BESS vendors must run qualification tests and submit data for NVIDIA review.

NVIDIA is not disclosing what, if anything, it charges for participation, and says more equipment categories will follow. Qualification is not deployment; read this as an ecosystem gatekeeping signal, not a capacity fact.

Baselayer (Osiris Ratings Inc.) announced on September 22 that it raised $35 million in Series A funding, led by M13 with participation from Torch Capital, Picus Ventures, Afore Capital and others. The valuation was not disclosed.

The funding was announced alongside a product launch: the Agentic Identity Suite, built on a "Know Your Agent" concept, which the company says cryptographically traces an AI agent's actions back to the deploying platform and the authorizing business. Baselayer says its existing business identity verification tools are used by more than 2,300 financial institutions, including one in five U.S. banks.

These are company statements relayed by SiliconANGLE; the round's completion status and the product's real-world effectiveness are independently unverified. As background for the market, Stripe said in June that more than 70% of requests to its API came from AI agents.

Hygon Information Technology launched its Hygon 1000 series embedded CPUs on September 22, targeting industrial automation devices such as robots and factory controllers, so computing and graphics processing can run locally.

The four-core, eight-thread processor is built on Hygon's x86-compatible C86 architecture and supports both DDR4 and DDR5 memory. It is the first time the Shanghai-listed company, previously focused on data centre CPUs, has extended its product line into physical AI devices.

The information comes from the vendor's announcement as reported by the South China Morning Post. Pricing, shipping timelines and customers were not disclosed, so actual adoption remains to be seen.

a16z co-founder Ben Horowitz and Udemy co-founder Gagan Biyani announced the Horowitz and Andreessen Academy on the September 22 episode of the a16z Show.

Per the episode's show notes, the school centers on students building real projects and working alongside companies and builders in San Francisco rather than studying theory first. It is positioned not as a college replacement but as a different path for young people who already know they want to build.

The notes say AI makes this an unusually powerful time to be young, but the school's size, tuition and admissions process are undisclosed and its ability to deliver is unverified.

The Biological Computing Company's "rat brain" AI model opens to select AWS customers on September 22 as a limited preview, after previously being available only through the small neocloud provider Bluesky Compute.

The technology uses activity patterns from rat neurons on multi-electrode arrays to improve video generation models. The company claims inference up to five times faster than an open-source video model it runs, but declines to name the comparison model. That is a vendor-tested figure with no third-party verification.

TBC has raised more than $50 million in total, including an additional $25 million first reported by WIRED. AWS executive Deap Ubhi himself raised the open question: whether fidelity holds when customers generate videos of ten minutes or longer remains unknown.

Kantata released Agent Studio on September 22. A user describes the job in plain language, and the platform assembles and refines a custom AI agent. The company says it is available now and in use by customers.

The tool sits inside the Kantata Expertise Engine, with the Expertise Agent launched in June doing the assembly. Every build and edit is checked for correctly structured inputs and outputs before publication, with no forms or code required. The intended users are practice leads and delivery directors.

The backdrop is Kantata's own diagnosis of the market: vendors are shipping narrow agents by the dozen, each scoped to a single job, and each arriving as another silo to maintain. Note that availability and customer usage are Kantata's own statements, not independently verified.

Members of the UN's International Scientific Panel on AI publicly pushed back against doomsday-style AI risk rhetoric at a forum held alongside the UN General Assembly in New York, with co-chair Yoshua Bengio urging a line between established science and reasonable extrapolation.

Joelle Barral, a Google DeepMind executive on the panel, said researchers should focus on clarifying what is settled and what remains unknown, and that stoking fear does not help. Nobel laureate Maria Ressa said the debate has swung between existential alarm and dismissing AI risk altogether.

Bengio acknowledged that experts themselves still disagree on the extrapolation side. This is an IT之家 report citing AFP coverage of the forum — statements by participants, not an official panel document.

Training only about 10 million trainable parameters (out of 598 million) on an already-trained SSL-based audio deepfake detector cuts pooled equal error rate from 4.85% to 3.74% across 14 cross-domain test sets.

Previously, improving cross-domain detection typically meant altering the original model's parameters or adding training data; CoReLoop leaves both untouched, reusing encoder outputs through lightweight refinement modules and low-rank adapters.

The 4.85%-to-3.74% pooled EER reduction is first-party, self-run by the authors; an optional halting head picks refinement depth per utterance, reaching 3.73% pooled EER at an average of 1.18 passes.

The composition of the 14 test sets and the identity of the baseline detector need checking in the full text, real-world generalization awaits independent replication, and the preprint was submitted on 2026-09-17.