Here is your Daily AI Stew, served up by Robin.
Nvidia's third-quarter revenue doubles to $115.8bn, and the board says it will not reinstate Huang
Start with tonight's most closely watched earnings report, which answers two questions at once: money and people.
According to Yahoo Finance, Nvidia reported third-quarter results with revenue up 103% year on year to $115.8bn, beating market expectations and raising guidance. This is Nvidia's first quarterly report since founder and long-serving chief executive Jensen Huang departed. By segment, revenue rose 103% year on year to $115.8bn, of which data-centre revenue rose 110% to $107.6bn. The board said it will not reinstate Huang as chief executive, and the company continues to be run by interim chief executive Cress.
What matters is not this quarter's beat but whether the company can hold its cadence after losing its long-time helmsman. Revenue rests almost entirely on the compute line, so customer concentration and order visibility will be the most important variables in the next report. By settling the interim arrangement as final, the board removes the personnel question in the short term and also leaves outsiders one less handle for watching strategic continuity. What to watch next is whether guidance is delivered, and whether gross margin gives ground once delivery costs rise.
Microsoft calls Copilot "the new operating system for work", with the Office trio built in
Over to the office line, where Microsoft gave its answer for the next entry point at its annual conference.
In the Ignite keynote, chief executive Satya Nadella laid out what the company calls its "frontier ecosystem" vision and announced that Word, Excel and PowerPoint are already formally available in Copilot. He called Copilot "the new operating system for work".
The framing inverts how things used to work: Word and its peers used to be the star, with the assistant a supporting player that helped you use it; Microsoft's vision is for agents to drive the work while software like Office drops back to the passenger seat. Apps become plug-in components inside agent workflows, with an organisational knowledge layer supplying context drawn from mail, meetings, calendars, chat and files. Users can create documents, spreadsheets and presentations directly in Copilot, and the generated files can still be edited on in the corresponding Office apps.
The wording deserves more attention than the features this time: Microsoft wants people to get used to starting from the assistant rather than from some application. Once the habit turns, whose models sit behind it no longer affects its position. The real threshold is on the enterprise side: that organisational knowledge layer has to connect to mail, meetings, calendars and chat, and until permissions, audit and liability for bad content are settled, procurement will not reach the signing stage. For teams building tools, the question changes too: the fight is no longer for users to open your app on their own, but to be called once by an agent at the right moment.
Google launches Gemini 4 Neon: a higher overall score, weaker long-horizon reasoning than Carbon
Google has launched Gemini 4 Neon, a new variant in the Gemini 4 series, which it officially positions as a high-capability model. The public material so far gives no official technical specifications, pricing, availability scope or verified benchmark results, so its real advantage over Carbon in the same series cannot be judged accurately yet.
Community observers, based on Artificial Analysis index data, noted that Neon's overall score is higher than Carbon's but that it scores lower on research-scale reasoning evaluations such as CritPt. That suggests Neon may lean more towards all-round capability, while Carbon still mainly targets long-horizon, complex research tasks. Other users say Neon reaches 127 tokens per second on OpenRouter against 46 per second for Carbon, which is a community observation that has not been independently verified.
In Artificial Analysis's Xhigh test, Neon's intelligence index is 61, above the median of 43 for comparable models; but this evaluation generated 283 million tokens, clearly above the median of 65.1 million, so production selection still has to work out token cost and response burden against real workloads.
Google uses names such as Neon and Carbon to distinguish the different models in the Gemini 4 series, but the frequent additions also make it hard for developers to judge which purpose and which capability difference each one corresponds to. Some commenters are weary of this ever-growing set of model names and of the lack of a clear mapping between names and uses, and some question whether the naming can tell you a model's scale tier; other users, judging by hands-on experience, think the latest Gemini is comparable to or slightly better than Opus 5.5 on some benchmarks, while its overall intelligent experience still falls short of the latter.
The credibility of such comparisons depends on test-set, latency and stability data, and those are exactly what has not been provided together. Engineering teams judge in a plain way: throw it at the hardest few changes in the repository, see whether they pass review, then work out the bill per million tokens and the time to first token. If it is to be used at all, canary it on a non-critical path first.
Microsoft ships seven in-house models at Ignite, with MAI Code 2 pitched at coding cost-effectiveness
Microsoft announced seven new AI models at Ignite 2026, covering tasks such as coding, image generation and editing, and transcription. Among them, MAI Code 2 targets code generation and is pitched on coding quality close to frontier models at a lower price.
Microsoft says MAI Code 2 can reach 98% of GPT-6.1 Sol's coding quality at about 40% of its price; against Opus 5.5 it can reach 96% of the performance at only 5% of the price. If these comparisons hold in real workloads, developers have one more cost-competitive code-model option, which could affect model choice and procurement. The available material, however, is mainly official framing and individual users' experience, lacking benchmark definitions, test methods, latency and stability data, and it does not set out the full specifications, open weights or licence terms of the seven models, so performance and cost-effectiveness still lack independent verification.
Microsoft cites Artificial Analysis rankings and price comparisons to show positioning, for example calling MAI Image 2.7 first in image editing, while MAI Image 2.7 Flash targets high-throughput generation; these are equally official launch framing, and the source provides no independent verification.
Some commenters take a positive view of MAI Code 2's hands-on coding experience, holding that its performance is comparable to Opus or Sol; others question whether the comparisons with GPT-6.1 Sol and Opus 5.5 are too bold, and press on why these models have not yet entered the Artificial Analysis index. Another view holds that Microsoft should offer open weights or an open-source version if it wants to expand market share; some commenters also criticise this Ignite launch as lacking innovation, saying they have no incentive to adopt Microsoft's AI stack.
The more names are spread out, the more the cost of choosing shifts onto users: several code names inside one series, no spec sheet from the vendor, and everyone can only guess from third-party leaderboards and benchmark scores. What deserves more attention is the evaluation basis — on the side with the higher intelligence index, the number of generated tokens is amplified dozens of times over, and the bill and latency are amplified with it. The real trade-off is not in a top-of-leaderboard ranking but in whether your task is the kind willing to trade time and money for reasoning depth. For everyday coding and writing alone, the premium on these high-scoring models may not be earned back.
Java4Shell: an OpenJDK Base64 decoding flaw can lead to remote code execution
The next one is engineering-heavy, but it concerns everyone running a Java service.
Calif disclosed an OpenJDK Base64 decoder vulnerability tracked as CVE-2026-114938. The root cause is that decoding can rewrite memory beyond the bounds of the target array, which creates the conditions for remote code execution. OpenJDK is the runtime basis of many Java services, and OpenJDK and Java SE deployments on affected branches face heap memory corruption, disclosure of confidential data and remote code execution triggered by an unauthenticated attacker, so the organisations concerned should inventory their exposure and patch promptly.
The disclosure also provides exploitation analysis, detection advice and mitigation guidance. The exploit requires heap-layout manipulation over the network, the so-called "Remote Heap Feng Shui", and its reliability is clearly affected by network round-trip latency and jitter, with low and stable latency favouring the attacker; the researchers completed a real attack test over a public IP in a controlled environment.
The current exploitation path is, however, strict about network conditions and the JVM environment, so the risk is concentrated on deployments that meet those conditions, and it cannot be treated directly as a generally and unconditionally exploitable vulnerability. Semgrep has already released a free detection script, and one of the five conditions it lists is that only HotSpot generates the relevant stub, so OpenJ9 and GCJ are unaffected. Commenters broadly recognise the complexity of this research and also think that, harsh as the conditions are, it may still affect a large number of hosts; the worry that critical infrastructure may soon be exploited for real is a personal judgment, while attributing the flaw to AVX512 Base64 decoding and working around it with -XX:UseAVX=2 is only speculation that has not been confirmed on social media.
What makes this weighty is not how easy the flaw is to exploit but how unevenly the patching is distributed: Java often sits inside someone else's product, operations may not even know which branch they are running, and the vendor sets the patch cadence. A detection script can help you screen out suspicious deployments, but after screening someone still has to confirm them one by one, and the window in between is where the risk sits. What is worth noting is the attack-surface line — moving a low-level trick like memory layout onto the network may cost reliability to latency, but it turns an attack that previously needed local privileges into one that is reachable remotely.
Google DeepMind's Genie 4 opens to some AI Ultra subscribers
One step further out, a model that generates interactive worlds in real time.
Google DeepMind announced that Genie 4 has launched and is available to eligible Google AI Ultra subscribers. Genie 4 is DeepMind's real-time generative world model, and the official rollout is staged: first to a limited set of trusted testers, then to eligible Google AI Ultra subscribers, with plans to open it to early partners in the Gemini API and AI Studio.
It builds on capabilities from the Gemini family, in which Gemini Omni makes the look and behaviour of simulations closer to the real world, and other Gemini models provide complex reasoning that helps keep simulations coherent and continuous. Eligible subscribers can use Genie 4 for real-time interactive world simulation, steering scenes with natural language while environmental sound, music and dialogue are generated in sync.
The available material, however, gives no specific capabilities, benchmark tests, performance data or technical improvements over the previous generation, and neither the access scope nor the subscription conditions are clearly stated, so some users cannot confirm whether their region or account qualifies, which limits the product's immediate impact.
Community discussion holds that Genie 4 is worth watching while questioning how thinly Google explains eligibility, region and the payment path; one user found that a workspace account could not subscribe, while a personal account page still showed Genie 3 as what Ultra offers. The discussion contains both appreciation for DeepMind's creative investment and for the possibility that it becomes a game engine like Roblox or GTA in future, and worries about insufficient profitability and potential negative social consequences, and someone also raised the technical question of whether scene detail stays consistent when the viewpoint rotates.
The value of a world model is not in demo clips but in whether it can be treated as an environment you can call repeatedly: teams working on embodied intelligence need to build thousands of edge cases in simulation, and games need scenes that stay consistent. What is stuck right now is not capability but availability — with eligibility, region and the payment path unclear, developers cannot put it into a plan. Another question nobody has answered is cost: generating a video-grade world in real time burns compute at every step, and whether a subscription can cover it depends on how it is priced once scaled.
Quad9 withdraws from France: public DNS choices under a court blocking order
Down from models to infrastructure, where public resolution services have seen a significant change.
Quad9 has withdrawn from the French market. Cloudflare's blog reads the event as one case of the growing tension between national regulation, public DNS operation and the governance of open internet infrastructure.
Quad9 is a free, non-profit public DNS resolution service that turns the domain names users visit into the corresponding IP addresses and provides network security protection. A French court ordered Quad9 to block sites accused of copyright infringement, and under this judicial obligation Quad9 announced that, from 24 November 2026, it will stop providing its free network protection service to users in France.
For French users, the withdrawal means one public DNS option fewer, and a need to turn to other privacy-friendly resolvers or to the resolution service their network operator provides. The case may also push other cross-border public DNS operators to adjust their services per national law and to exit specific markets when compliance requirements conflict.
Quad9's chief technology officer said the IETF 127 meeting at the time was 100% behind its position, but stressed that this was only a point-in-time consensus, and described widespread concern among attendees about the situation DNS and other areas are facing. Most other commenters criticise the internet being over-regulated, while one UK user holds that the French case is not necessarily the proverbial canary in the coal mine and may simply be an example of a national government successfully implementing internet regulation.
This is not one organisation's commercial choice but a structural problem when public infrastructure runs into national jurisdictions: resolution is free, yet the compliance cost is concentrated on a few operators, and once requirements across countries conflict, withdrawal becomes the least troublesome option among the answers. The consequences for users are concrete: swap the resolver and privacy protection and filtering capability change with it, and nobody usually comes to tell you. What to watch next is whether similar operators follow, and whether local alternatives can fill the gap.
Will Wright's new simulation game MyMems mines player memories with generative AI
One more, a lighter topic, though it hides a new direction for putting AI into practice.
Will Wright, designer of SimCity and The Sims, had been quiet for a long time since Spore shipped in 2008 and he left EA, which Maxis belonged to. Now he has started promoting his next digital simulation work: a product that uses generative AI to generate the game experience from players' memories.
The game is called MyMems, and it is planned for release next year. According to the press release from Wright's Gallium Studios and partner Artificial Agency, it turns simulation inward — the game mines memories supplied by players and generates digital characters, which can form teams and solve puzzles in the game world.
In the press release Wright gives two examples: "If my grandmother played a Grand Slam match against Serena Williams, who would win?" and "If I had to build a bridge over a not-so-wide ditch, which three people from my own life would I pick to help?"
What is interesting is that the direction is reversed: simulation games used to hand you a world to walk into, and this time your own experiences are the material, for the system to arrange. The ceiling on this kind of approach is not generation quality but the sensitivity of memory itself — how much private content players hand over, where the data is stored, and who has seen it decide whether ordinary households can accept it. The other variable is whether the gameplay holds up: putting people you know into a scene to solve puzzles is novel the first time, and by the tenth time it comes down to whether the system really understands those people.
Put today's items side by side, and the entry point is folding back towards the assistant layer, models are being tiered by capability and price, and even public infrastructure has started picking its markets. The technical bill is easy to compute; the bill for responsibility and cost still has to be settled by developers themselves. That is all for today.



