This is the daily AI stew, brought to you by Robin.
Nvidia announces a CEO handover: Jensen Huang steps down immediately, stock falls 14%
Nvidia moves to replace CEO Jensen Huang, an OpenAI Foundation director warns of an AI takeover within months, and the FAA expands its AI air-traffic advice to 22 centers.
Nvidia has begun a CEO transition: long-serving chief executive Jensen Huang will leave the company and its board, effective immediately. During his tenure he turned Nvidia into the leading AI chip company. Axios reports the company has started looking for a permanent successor and lists 11 possible candidates. Public information so far names no confirmed successor, gives no handover timetable, and quantifies no financial impact. After the news took effect the stock fell as much as 14%, a sign investors are repricing the leadership change in real time.
When a company's valuation rests on one person for years, a personnel change becomes a risk event, not just headline noise. The market's immediate reaction shows how heavily investors still weight the founder. The next thing to watch is whether the company gives a succession plan and timetable soon; a transition with no timetable often costs more than the transition itself. The successor also affects the technology roadmap and external relationships, so this is no longer purely a personnel matter.
OpenAI Foundation director warns AI could "take over" society in months
Someone in the industry put it in the starkest terms, and this one deserves a full listen.
OpenAI Foundation director and Alignment Research Center executive director Paul Christiano, speaking this week at a lecture series at the Toronto Convention Centre, listed AI accidents among the leading causes of death. In the first of three lectures he opened by saying that for most people in the room, the most likely cause of death would be an AI accident. He defined the "AI takeover" he fears as a state in which goal-directed AI systems "give orders like humans do today" and decide what happens on Earth, while anyone trying to stop them "would be as ineffective as you or I trying to stop the US military." Christiano said such a situation would very likely bring mass death, and could even make the environment unfit for humans.
Treating a practitioner's public lecture as prophecy is not right, but it does show that this circle's sense of the pace of risk is quickening. The real dispute is the timescale: if the consequences land in years rather than decades, the priorities for regulation and engineering practice change completely. Worth watching is whether research institutes and foundations turn statements like this into concrete evaluation standards or funding directions.
FAA expands AI traffic advice to 22 air-traffic centers nationwide
On the institutional side there is also a concrete rollout, and a very specific one.
The US Federal Aviation Administration has expanded an AI tool that advises air-traffic controllers from the capital region to all US airspace. In a 10 November statement the FAA said the system, called SMART, cost $875 million and has helped reroute 180 flights since going live over Washington DC airspace on 21 September, saving about 10 minutes and nearly 80 gallons of fuel per flight on average. The statement said that after the capital-region pilot succeeded, SMART will be used across the airspace of 22 control centers nationwide to cut delays and fuel burn. Airline executives backed the expansion, but the US union representing controllers said its members took no part in designing or testing the system.
An advisory system can make a controller's job easier and make a union nervous at the same time. The disagreement is whether it advises or replaces judgement, and who is accountable when something goes wrong. The spending shows regulators are willing to pay for efficiency, but the performance numbers still come from the side that deployed it. Watch whether the union is brought into later validation and training, and whether records of rejected advice are published.
Google's Gemini smart glasses go on sale from $395, with privacy as the pitch
Back to consumer hardware: this product went the long way round and is back on stage.
Google has put Gemini into a pair of glasses and opened preorders from $395. The devices are a collaboration between Google, Samsung and two eyewear brands. Warby Parker designed three frames, Alden, Reid and Wren, in Pine, Saddle and Jet Black, with standard or sun lenses, from $395, sold only in the US. Gentle Monster brought six styles in deep red, pink, grey and yellow, with clear, tinted, blue or brown lenses, from $449, sold in the US and Korea. Samsung supplies the computing hardware in the temples and Google supplies the software. Smart glasses have been dogged by privacy worries since the Google Glass era, and this time Google is leading with privacy.
Making privacy the selling point is an admission that the first hurdle for this class of device is not technology but trust. Going through two eyewear brands' stores is far faster than building retail, and it spreads the design and brand risk. What matters later is not unit sales but whether the recording indicator is obvious enough, and how much data third-party apps actually get.
Google releases Gemini 4 Hydrogen: a model scheduler for coding agents
Changing subject, here is what moved at the model layer today.
Google released Gemini 4 Hydrogen, alongside a Gemini 4 Flash model and a scheduler that coordinates several Gemini models inside agentic coding workflows. The idea is to let an agent dispatch work by task difficulty and capability: routine edits to a faster model, hard reasoning to a stronger one, which cuts manual switching. Google lists routing, caching, latency and cost as the four core design variables. One number cited in discussion: with co-trained models and a dedicated execution framework, about 25% can be saved; another claim says Mjolnir saves 60% by wiring Opus to DeepSeek Flash over ACP, but that is an anecdotal observation. Commenters generally see this as a continuation of the advisor pattern and multi-model pipelines rather than a new category, with the real change in integration and ease of use. For developers running multi-model coding agents, Hydrogen adds a fast, cheap execution layer, leaving Carbon and Argon for long, deep-reasoning tasks. Commenters also question whether the cache can be reused, whether the model should choose its own route, and the real latency and savings; some report Argon and Carbon feel slower than 3.8 Flash and Opus. Public material gives too little implementation detail and too few benchmarks, so the 25% figure is not independently verified.
Putting routing into the product means cost control moves from an engineering detail to a vendor-provided default: teams write less glue and give up some tuning space. What to watch is whether the cache can be reused across calls, and who makes the routing decision; handing it to the model makes failures hard to attribute. The savings quoted in public material come from the vendor and community observation, so waiting for third-party runs before putting them in a budget is the safer bet.
llama.cpp ships 1.0: local inference gets a semantic version contract
Back to developers' desks: this is an old project's coming of age.
The open-source local inference project llama.cpp released version 1.0. The maintainers stressed two things: continued support for existing platforms and model types, and a lean, solid base for downstream applications. Alongside 1.0 they announced the Llama desktop app built on llama.cpp, aimed at making local models a few clicks away. For downstream developers the most practical part of 1.0 is the version contract: the public C API now follows semantic versioning, so a major bump may break API and ABI, a minor bump only adds backward-compatible features, and a patch only fixes bugs without touching the API. With those two desktop front ends, the setup needed to get local AI running loses another layer.
For a project countless applications build on, making the version number explicit is worth more than adding features: downstream packagers and distribution maintainers can now predict whether an upgrade will break. The desktop app moves the first mile from the command line into a GUI, which changes the user base and the kind of bug reports that arrive. Next, watch whether it holds compatibility under the version contract; that decides whether it moves from hobbyist tool to real dependency.
Huawei open-sources MACaQ quantisation, with a llama.cpp fork and benchmarks
Another story in the same lane landed the same day, from Huawei.
Huawei open-sourced a quantisation method called MACaQ, together with GGUF support for llama.cpp, a code fork, evaluation scripts and a paper. By the project's own measurements, on Qwen3-4B at IQ4_XS the MACaQ file is 2.260 GB with an average KL divergence of 0.02804, about 18.6% below Bartowski's 0.03443 and about 23.5% below Unsloth's 0.03643, at broadly the same file size. On Qwen3.6-35B-A3B the average KL divergence is 0.018696, below Unsloth UD-IQ4_XS at 0.020954 and Bartowski at 0.021927. On Qwen3.8-27B the figure is 0.036784, about 65% below Bartowski's 0.104170, though the file sizes compared are not exactly equal. Format support currently covers 2-bit IQ2_XXS and IQ2_XS, 3-bit IQ3_XXS and IQ3_S, and 4-bit IQ4_NL and IQ4_XS. The project also claims that 32.5 GB of VRAM can quantise a 122-billion-parameter model in 64 minutes. All these quality and efficiency figures come from the project's own tests; no third-party reproduction has appeared.
Quantisation comparisons have always been hard-core in the local model scene, because a few percentage points of file size decide whether a card can run a model at all. A vendor self-test beating existing options is not surprising; the hard part is reproducing it with the same model and the same quant level. What really decides adoption is whether the GGUF fork can keep pace with llama.cpp mainline, and whether the evaluation scripts are simple enough that third parties bother.
Zyphra open-sources ZONOS3-Mini: 800M parameters, TTS in 43 languages
The voice line has an update too, small in scale but clear in aim.
Zyphra released a compact open-source text-to-speech model, ZONOS3-Mini, with weights under Apache 2.0. It uses a latent MoE architecture with 64 experts and top-2 routing, 800M total parameters but only about 100M active per token, which is the key to running on-device. The model supports 43 languages, can clone a voice from up to 60 seconds of reference audio, and can adjust age, accent, timbre, emotion, speaking rate and per-word expression; the core acoustic controls are three buckets for loudness, bandwidth and rate. Runtime support covers NVIDIA CUDA, Apple Silicon MLX, C++/ggml CPU/CUDA/Metal/Vulkan, and browser WebGPU; a unified HTTP API can load GGUF weights, WebGPU shards and 27 default voices, and the hosted Zyphra Cloud version can be used directly. It uses a 32,768 context length and can synthesise up to 10 minutes of 44.1 kHz audio in one pass. There is no independent third-party evaluation or per-device benchmark yet, so cross-language quality and real-time performance still need measurement.
Pushing TTS to the edge changes latency and privacy first: audio need not leave the device and there is no per-call fee. Many languages plus controllable timbre and emotion effectively turn voice-over work into parameter tuning, which is directly attractive to podcasters and short-video makers. But once voice cloning spreads, licensing and impersonation will become a problem before the technology does, and independent comparisons of cross-language quality are still missing.
Voltage glitching unlocks the CMP 170HX: a mining card becomes an 80GB inference accelerator
To close, a wild one from the hardware scene.
Someone used voltage fault injection to unlock Nvidia's mining-oriented CMP 170HX into something close to an A100, with up to 80GB of usable HBM. According to the author, Kate, the modification enables 120 to 126 streaming multiprocessors against 108 on a retail A100, and restores NVLink and PCIe Gen4 support. FP16 measured 85.9 TFLOP/s against 72.5 TFLOP/s on a retail A100, with all 80GB usable for inference. In principle the CMP 170HX uses an A100-class die, with most HBM and the extra SMs disabled at the factory; software can unlock some memory, but deeper hardware fuse locks need voltage glitching, briefly disturbing the card's power behaviour at boot. The project therefore reconfigures existing hardware rather than adding memory or compute from nothing. The cost is real: board rewiring, a home-made STM32 glitcher, and 1 to 48 hours of glitching per card; one of the author's ten cards ended up with only 64GB usable. Demand pushed prices up too, from roughly $100 to $200 for a modded card to $1,200 to $2,000.
The value of this kind of mod is not showing off but reminding the market that compute locked behind software and fuses is itself a grey secondary supply. A tenfold price rise shows the demand is real, and that the window is short: one tightening by the vendor or a dry second-hand supply and it closes. For small teams the saving has to be weighed with the risk; there is no stability or warranty, and one card in ten came back with reduced memory, by the author's own account.
Today's news has people and org charts moving on one side, and the toolchain moving into finer layers on the other: model routing, quantisation, voice, even turning mining cards back into inference. That is all for now; we will see what is new tomorrow.
