Today's AI Pulse
OpenAI unveils ChatGPT Images 2.0 with multilingual text and ‘Thinking’ mode
OpenAI rolled out ChatGPT Images 2.0, a next‑generation image model that adds multilingual typography, real‑time web research, and agentic reasoning. The new ‘Thinking’ mode lets paid users generate up to eight coherent images per prompt, including floor plans, maps and UI mock‑ups. The feature expands the model’s ability to work with uploaded content.
Also developing:
Perplexity Says Its AI Personal Shopper ‘Puts You First’
Perplexity has launched a free AI‑powered personal shopper for U.S. users ahead of the holiday season, allowing shoppers to type queries, refine results and purchase items directly through the platform via PayPal’s Instant Buy integration. The assistant remembers prior interactions to tailor recommendations, displays product cards with specs and reviews, and is currently available on desktop and web with iOS and Android apps arriving in the coming weeks. Perplexity positions the service as a more intent‑driven alternative to traditional search bars and affiliate‑heavy editorial sites, emphasizing a "joy of shopping" experience rather than just fast checkout.
Dev Roles Shift From Coding to Code Evaluation
How do devs feel about this job change? Example below in Codex. Pros: you can kick off tasks from anywhere (just talked to one dev who started multiple codex tasks while getting into a cab), you get multiple versions to pull...
TPU Scarcity and Memory Lag Behind A100s, Prompting Rework
@_The_Prophet__ TPUs had low availability for ages and also low memory relatively on the v6e especially versus the hoppers working pretty much out of the box similar to a100s Grace Blackwell is the next thing that needs reworking so there is...
The Genesis Mission's Impact on the Global AI Race
President Donald Trump announced the Genesis Mission, a DOE‑led initiative that unites 17 federal labs with tech giants such as Nvidia, Microsoft, and OpenAI to fuse supercomputing, AI and quantum capabilities. The platform aims to generate massive scientific datasets, accelerating...
Must‑Listen Interview Ignites AI Community, Follow My X Pro Lists
I agree. Incredible interview by @dwarkesh_sp of @ilyasut. I could listen to both for months and not get bored. It's like being at a great university and hearing the best professor. I love X. This just LIT UP the AI community....
AI Judges Will Filter the Coming Content Deluge
One AI use case that is only getting more popular: LLM as a judge. Everyone still talks about AI generating more content, but not enough people are talking about: 1) the horrific deluge of noise we're going to have to deal...
Wyze’s New Security Camera Watches Your Yard From Inside Your Home
Wyze has launched a $34.99 Window Cam that monitors a yard from inside a rear window, eliminating the need for batteries, exterior power, Wi‑Fi extenders, or weatherproofing. The 1080p camera offers a 101-degree horizontal field of view, enhanced color night...
Scaling Boosts Benchmarks, Not Genuine Problem‑solving Ability
I think it is somewhat true though that scaling helps with benchmark performance but not necessarily with with new model capabilities. Like the example he mentioned > U: "Please code xyz." > M: "Ok here is xyz." > U: "You have a bug." >...
AI Control Remains Unsolved; Top Oversight Fails 92%
Excited to present our new AI paper as a @NeurIPSConf spotlight next week: we find that the problem of controlling artificial superintelligence remains unsolved. With simulations and scaling laws, we find that an implementation of the least unpromising...
After LLMs: Spatial Intelligence and World Models — Fei-Fei Li & Justin Johnson, World Labs
Fei-Fei Li and Justin Johnson discuss their new platform Marble, a generative world model that turns text, images, and spatial inputs into editable 3D environments, highlighting its technical core of Gaussian splats and real‑time interactivity across devices. They argue that...
OpenAI Positions Codex as Team Member, Not Tool
OpenAI is very deliberate about how they talk about Codex. It's not positioned as an operating system. It's heavily positioned as a teammate. Their site says: "Your new coding partner", "accelerates your team" Their job postings say: "we're building an AI software...
SAM 3D Enables Data‑Driven, Personalized Rehabilitation Insights
SAM 3D is helping advance the future of rehabilitation. See how researchers at @carnegiemellon are using SAM 3D to capture and analyze human movement in clinical settings, opening the doors to personalized, data-driven insights in the recovery process. 🔗 Learn more about...
The Trump Administration Just Launched Its Own Plan for Global AI Dominance and What Could Go Wrong?
The Trump White House issued the Genesis Mission Executive Order, directing the Department of Energy to create a government‑run AI platform – the American Science and Security Platform – that will leverage federal scientific datasets to train foundation models and...
Open Collaboration Fuels AI Boom and Scientific Innovation
Excited about the Genesis mission - congrats to @POTUS @SecretaryWright @ScienceUnderSec @mkratsios47 @sriramk! We've experienced first-hand how more openness and collaboration in the US can massively accelerate progress. In my opinion, that's what led to the current AI boom and US...
Adaptive 2D Gaussians Redefine Image Compression and Restoration
📢 Image-GS: Content-Adaptive Image Reconstruction using 2D Gaussians In this week’s deep dive, we explore Image-GS, a groundbreaking framework that reimagines how images can be represented, compressed, restored, and upsampled using adaptive 2D Gaussian splats. Unlike traditional codecs or neural...
I Found the World's Fastest Mini PC with an Intel CPU that Trounces the Mac Pro's M2 Ultra — and...
Lenovo’s ThinkCentre M90s Gen 6 mini PC, equipped with an Intel Core Ultra 9 285 (24‑core, 65 W) processor, outperforms AMD’s Ryzen AI Max+ and Apple’s M2 Ultra on PassMark benchmarks while fitting into a 340 × 93 × 300 mm chassis. Priced at $1,329 after a $570 Black‑Friday...
New AI Tests Need Fresh Images; Recipe Finally Clarified
One more comment is that giving this image to an AI and asking about it is not sufficient to show the diff because it's all over the training data by now. You'd have to use a new, very recent image,...
Pretrain, Fine‑tune, and Let Big AI Solve Tasks
@matejhladky_dev AI has crushed it since this post way beyond expectation. I made the same category of mistake all of AI was making, of thinking we have to discover and write the algorithm. You don't. You pretrain and then finetune...
Solar-Powered iLamp Turns the Humble Lamppost Into an AI Hub
British greentech firm Conflow Power Group has launched the iLamp, a solar‑powered streetlight that doubles as a micro AI data centre using Nvidia Jetson processors. Each unit generates 200‑600 watts from a self‑cleaning panel, consumes 80 watts for lighting and...
LLMs Know Popular APIs, Need Docs for Obscure Ones
I've had medium success asking LLMs if a thing exists, it works out of the box for some of the more well-known things (e.g. both GPT 5.1 and Gemini 3 know about this function if you describe the tensor transformation...
Discover PyTorch’s Pixel_unshuffle: Skip Custom Tensor Hacks
Always a slightly mixed feeling to write pretty good first-principles code to do some tensor rearrangement, only to find that PyTorch has a built in function that does it faster. I had made a point of at least skimming the docs...
What Enterprises Should Know About The White House's New AI 'Manhattan Project' The Genesis Mission
President Trump announced the Genesis Mission, an AI‑focused "Manhattan Project" that directs the Department of Energy to create a closed‑loop AI experimentation platform linking the nation’s 17 national labs, federal supercomputers and decades of government scientific data. The initiative aims...
Clear AI Rules Accelerate Trust and Innovation
85% of organizations believe responsible AI is a top management issue. Yet only 25% have governance mechanisms in place to address it. This trust gap is costing companies dearly. In Europe alone, 68% of companies don't understand their EU AI...
AI's Temporal Hacking Threatens Autonomy and Democracy
AI manipulation techniques revealed. For our free newsletter this week, we cover temporal hacking: AI systems that game human attention over months. @IrenaCronin and I write this newsletter every week. Temporal hacking describes AI systems that optimize for long term outcomes by subtly...
AI Landing Page Tools Score Poorly in Real Audits
We scored every major AI landing page analyzer across the same criteria we use for real CRO audits. Comprehensiveness. Specificity. Originality. Realistic implementation. Correctness. The highest score was 5/15. Several landed at zero. This isn’t a knock on AI. It’s a reflection of...
Chat UI Limits AI; Context‑First Unlocks Power
Chat made AI feel real for the first time. A blank box. A question. A response that sounded alive. It created the belief that conversation was the natural interface for intelligence. But the chat window is the smallest view of...
Chat UI Masks AI Limits; Context Drives True Potential
Chat made AI feel real for the first time. A blank box. A question. A response that sounded alive. It created the belief that conversation was the natural interface for intelligence. But the chat window is the smallest view of...
Scaling Pre‑training Hits Diminishing Returns for Future Generations
@GiorgioMantova @dwarkesh_sp @ilyasut I’d say this is the jump from last gen to current gen, but I think the argument is that further improvements will fizzle out in the next gen if we keep scaling pre-training. Ie it won’t give...
Flux.1-dev Ranks #2, Eagerly Awaiting Flux.2-dev
Flux.1-dev has been the second most liked model on Hugging Face just after Deepseek R1 so super excited to see the release of Flux.2-dev by @bfl_ml today! Download the weights or try the model (thanks to @fal) on @huggingface: https://t.co/kdmVlvdLZh Read the...
Reachy Mini Becomes My New Podcast Assistant
Reachy mini is my new podcast assistant! Coming soon with @ti_morse... https://t.co/VfUQn1Cgz6
Grok Imagine Impresses with Quality Hindi Generation
@DevDminGod this is pretty good! i didn't know grok imagine could do such decent hindi
TPUs Offer More Stable Training Than CUDA on Large Batches
@_The_Prophet__ TPUs have been more stable for training than CUDA equivalents for a couple of years now, especially on large batch sizes XLA is pretty good now! For inference it makes even less of a difference (We previously trained sota models on thousands...
Image Generation Has Leaped Forward Since Summer 2022
it is wild how far we’ve come since the hot image gen summer of 2022 image cred: @bfl_ml https://t.co/i5KbszlFDM
FLUX.2 Launches, Upgrading Community’s Top Image Model
the community’s favorite image creation and editing model just got better: welcome, FLUX.2 by @bfl_ml 🤩 https://t.co/iLrbYYK4bd
CV Engineers Should Master YOLO for Object Detection
So you are a CV engineer, what do you know about Computer Vision? I have used YOLO for... https://t.co/lX4OrFFYqE
New Research Paper Now Available Online
Here's our paper: https://t.co/RmNft3zU5Z
Beyond Scaling: Engineering Tricks Now Drive AI Progress
@dwarkesh_sp @ilyasut “The Age of Scaling is over.” I agree with that. Basically, since GPT 4.5 a lot of the perceived real-world progress was driven by clever engineering wrappers (context filtering, inference scaling, multi-turn tricks, retrieval, tool use, etc).
15 Essential Architectural Traits for Building Robust AI Agents
Just shared this brilliant mind map on the 15 key architectural characteristics of AI agents — absolutely packed with insights! Modularity, evolvability, context awareness, security compliance… everything you need to design robust agents. Huge thanks to @Python_Dv for creating this gem
Seeing Benchmaxxing, Ilya Launches Company for Proper LLM Development
Ok, so what Ilya saw was extreme benchmaxxing, which in turn prompted him to create his own company to do LLM development the proper way?! Makes sense, I sympathize with that.
Machines Exploit Shortcuts, Creating More Correct‑unintended Rules than Humans
@giffmana @dileeplearning the "correct-unintended" rules were just that -- correct on the demonstrations but using "shortcuts" (e.g., the numerical value of a color). We also saw a small percentage of "correct-unintended" rules that humans generated, but much less...
Machines Craft Meaningful Unintended Rules; Humans Produce Nonsense
@giffmana @dileeplearning There was a big difference between "not classified" rules generated by humans and "correct-unintended" rules generated by machines. For humans, the "not classified" rules were generally humans writing nonsensical things like ⬇️
AI and Energy Tech Converge to Transform Industries
@UmmayHabiba0 @SchneiderNA Certainly. We’re witnessing a major shift in real time. AI and energy tech are finally converging in ways that will reshape how industries operate and how infrastructure is built. Here’s the video if you’d like to take a look: 📺...
AI, Energy, Infrastructure Converge to Transform U.S. Economy
@the_AI_girl @SchneiderNA Absolutely, the momentum building across AI, energy, and infrastructure is setting the stage for a major transformation in the U.S. economy. I just shared more of my insights here on @LinkedIn : https://t.co/WwaOkGdcNm Big shifts ahead.
Hybrid Search on 1.2M Samples: BM25 & Embeddings
1.2 million samples. BM25, Embeddings and Hybrid search. Tutorial and code comes tomorrow! Stay tuned! https://t.co/FlmaDlpASR
Embeddings Beat LLMs for Fast, Cheap Classification
totally forgot about this experiment where i found it was faster and cheaper to do classification via embeddings vs using the fastest/cheapest llm (at the time)
Embedding Classification Beats Fastest Model on Speed, Cost
@jasonth0 did an experiment a bit back, and found that embedding based classification seemed consistently faster and cheaper than using the cheapest/fastest model (at the time) https://t.co/uuEPwu88cg
LLM
kinda like this, but instead of using vec2text - i found grabbing a few samples from each cluster and feeding into an llm came up with better names (not surprisingly) https://t.co/K8phyyDFdR
Clustering Embeddings Drives Dynamic Ontology Exploration
last weekend i went down the rabbit hole of how to build dynamic ontologies, and kept coming back to clustering of embeddings curious if anyone has cool experiments i could look at around this
Anthropic Takes Lead in AI Coding Competition
Anthropic won the AI coding race 😏
Modern VLMs Struggle with Long‑Horizon Household Tasks
Our most recent work that benchmarks modern VLM and their efficacy for long horizon household activities in robotic learning, using BEHAVIOR benchmark environment.👇