Gemini 3 arrives as OpenAI attempts to combat it with a new Codex release
Get bigger weekly updates! For a limited time, you can grab a month-long trial of our paid newsletter which includes an extended weekly updates and some other benefits to keep up with the rapid AI news cycle.
last week’s top stories
🚀 Google launches Gemini 3 AI model. Google introduced Gemini 3, the latest version of its general AI, immediately integrating it into Google Search and other products. Sundar Pichai calls it their “most intelligent model,” with top-ranking performance on industry benchmarks and improved reasoning & coding skills.
🎨 Google rolls out Nano Banana 2 image AI. Google DeepMind unveiled Nano Banana 2 (aka “Nano Banana Pro”), a new text-to-image model built on the Gemini 3 AI. It generates higher-resolution images (up to 4K) with sharply rendered text and can even search the web for real-time info to include in visuals. Aimed at designers, it offers fine control over image details (camera angles, lighting, etc.), though its high fidelity outputs are slower and costlier than the previous version. Read more
💻 OpenAI debuts GPT-5 Codex-Max coding model. OpenAI launched an upgraded code-focused AI, GPT-5.1-Codex-Max, built to autonomously tackle large coding projects by “thinking” much longer and compacting its context. The model can work for hours, summarizing and pruning context to handle multi-million-token tasks (it even completed a 24-hour refactoring job internally). It outperforms previous Codex versions on coding benchmarks (77.9% on a bug-fix test vs 73.7% prior) and now replaces GPT-5.1 as the default code engine for ChatGPT’s developer tools. Read more
🔮 Anthropic’s Claude Opus 4.5 nears release. Rumors swirl that Anthropic is poised to launch Claude Opus 4.5, an advanced “frontier” version of its AI model. Leaked info spotted on an AI platform suggests the model (code-named “Kayak”) could arrive as early as today. Opus 4.5 would follow Anthropic’s recent Claude 4.5 “Sonnet” and “Haiku” models, and is expected to further boost coding prowess and autonomous reasoning. Read more
👥 ChatGPT adds group chat feature. OpenAI expanded ChatGPT to support shared conversations among multiple users. The new group chats (up to 20 people per room) let friends or co-workers collaborate with each other and ChatGPT in the same thread. ChatGPT can intelligently join the discussion when tagged, helping summarize or provide answers without dominating the chat. Read more
🦾 xAI launches Grok 4.1 model. Grok 4.1 topped several public leaderboards, briefly holding the #1 spot on LMArena’s test before Google’s Gemini 3 came out. It features two modes: a fast mode for instant replies and a “thinking” mode that does deeper reasoning in steps. Under the hood, Grok 4.1 greatly improved multimodal abilities (it can analyze images/videos and handle 1+ million token context without losing coherence) and slashed its hallucination rate by ~65% versus the previous version. Read more
⚠️ Anthropic CEO issues new AI warnings. Anthropic’s chief Dario Amodei voiced deep unease about AI’s trajectory in a 60 Minutes interview. He said he’s “deeply uncomfortable” that a few tech CEOs (himself included) are making unilateral decisions about AI’s future. Amodei warned that advanced AI could outsmart humans and even wipe out 50% of entry-level white-collar jobs within 5 years without proper regulation. Read more
🫧 Google CEO warns of AI investment bubble. Alphabet’s Sundar Pichai cautioned that today’s AI funding frenzy shows “elements of irrationality” akin to the dot-com era. In a BBC interview, he noted soaring valuations and heavy AI R&D spending could prove unsustainable and said “no company is going to be immune” if an AI bubble bursts (including Google itself). Google’s stock has boomed on AI optimism, but Pichai urged realism about challenges like AI’s immense energy costs, which he warned are so high that Google will delay some climate goals while scaling AI computing. Read more
🗣️ Meta open-sources speech model for 1,600 languages. Meta released an Omnilingual AI for speech recognition that natively transcribes 1,100+ more languages than any previous model. Dubbed “Massively Multilingual Speech” (MMS), it covers 1,600+ languages (vs. 99 in OpenAI’s Whisper) and can even generalize to 5,000+ languages via few-shot learning. Meta open-sourced the models under Apache 2.0, letting developers use them freely even commercially. The suite includes a 7B-parameter wav2vec 2.0 encoder and several ASR decoders, plus a giant speech dataset of 3.5M hours. Read more
⚖️ UK court sides with Stability AI in Getty case. In a landmark ruling, a London high court mostly rejected Getty Images’ copyright claims against Stability AI. The judge held that Stability’s image generator (Stable Diffusion) does not store or reproduce the original training images, so the AI model itself is not an “infringing copy” of Getty’s photos. Getty had sued over Stable Diffusion being trained on millions of internet images (some Getty-owned), but key parts of the case were withdrawn or dismissed due to jurisdiction and technical nuances. Read more
📈 Nvidia shatters records on AI chip demand. The chipmaker Nvidia reported astronomical earnings, quelling talk of an “AI bubble” for now. Q3 revenue hit $57 billion (up 62% year-on-year) with $32 billion in profit, blowing past forecasts. Demand for Nvidia’s AI GPUs is so high that data center sales jumped 66% to $51B in the quarter, and CEO Jensen Huang said orders for ~$500B of chips (through 2026) are already in the pipeline. “We’ve entered the virtuous cycle of AI,” Huang noted, as cloud providers and enterprises worldwide race to buy Nvidia silicon. Read more
🧪 AI Research of the Week
The cost of thinking is similar between large reasoning models and humans
From MIT McGovern Institute
Jake’s Take: The study measured how much thinking budget humans and modern reasoning LLMs spent on the same puzzles and found a clear match between human response times and the number of internal tokens the models consumed.
The team gave seven classes of problems, from arithmetic to ARC grid puzzles, to both humans and chain of thought tuned models, logged millisecond reaction times for people and hidden reasoning tokens for the models, and saw that items that delayed humans led models to grow long internal chains.
This result suggests that models being built share a meaningful similarity with how humans reason, with a similar level of effort and problem difficulty. Do you think your reasoning skills are better than a bot’s?
and then, even more news…
🇪🇺 Europe warned it’s lagging in AI race. ECB President Christine Lagarde warned that Europe is “missing the boat” on AI and risking its future. In a Nov 24 speech, she said the EU has “already missed the opportunity to be a first mover” in AI as the US and China pour investment into the field. Lagarde urged Europe to remove barriers and adopt AI faster, noting that unlike past tech waves, AI’s economic payoff (and disruption) will come much sooner. Read more