OpenAI●Covered by 9 sources
OpenAI built its own chip in 9 months and named it Jalapeño
OpenAI and Broadcom unveiled Jalapeño, a custom inference chip designed to run ChatGPT and Codex with much better performance per watt, already running GPT-5.3-Codex-Spark in the lab with full deployment slated for end of 2026. Community teardowns peg it at a TPU-like design with roughly 7.1–7.4 TB/s bandwidth and ~10 PFLOPS FP4, and OpenAI used its own AI models to speed up the design.
So what? ↓Every dollar OpenAI stops paying Nvidia for inference is a dollar it can stop charging you, so the point of homemade silicon is making ChatGPT cheaper to run and harder for one supplier to bottleneck.
Source: OpenAI
theverge.com●Covered by 6 sources
Apple just raised Mac and iPad prices 15-25%, and it's blaming the AI memory crunch
Macs, MacBooks, iPads, smart speakers, and even the $599 MacBook Neo (now $699) got more expensive today, with some products jumping by as much as $500. Apple says it has "never seen a component price increase this much, this quickly," pinning it on AI data centers hoovering up the world's RAM and SSDs; the iPhone is untouched for now.
So what? ↓The data-center gold rush you keep reading about just showed up on the price tag of the laptop you were about to buy, which is the most direct way the AI boom has reached into a normal person's wallet so far.
Source: theverge.com
Google●Covered by 5 sources
Gemini 3.5 Flash now scores 78.4 on screen control, dead even with GPT-5.5
Google baked native computer-use into its lightweight Gemini 3.5 Flash model, letting it read continuous screenshots and click, scroll, and type its way through browsers, apps, and desktops. The OSWorld score ties GPT-5.5, and developers can wire it up through the Gemini API for software testing and office automation.
So what? ↓This is the cheap-and-fast model, not the flagship, so the moment that can drive your actual screen as well as the expensive one, the desk-job automation that used to be a premium demo gets priced for everyone.
Source: Google
TechCrunch●Covered by 4 sources
At least two Gemini researchers walk out of Google's door straight into Anthropic's
Jonas Adler and Alexander Pritzel are the latest DeepMind names to leave Google for Anthropic (one report adds a third, Arthur Conmy), following earlier exits by Noam Shazeer and John Jumper. The trickle of senior Gemini and DeepMind talent toward rivals is now a steady stream.
So what? ↓The people who build the models you'll be paying for keep choosing a competitor, and where the brains go, the better products tend to follow.
Source: TechCrunch
engadget.com●Covered by 3 sources
IBM crammed 100 billion transistors onto a fingernail and called it sub-1 nanometer
IBM says its new prototype "nanostack" chip doubles the transistor density of its 2021 state-of-the-art design, packing roughly 100 billion onto a fingernail-sized area, which it pitches as a path to keep Moore's Law limping along another decade. It's a lab prototype, not something shipping in your laptop.
So what? ↓More transistors per chip is the unglamorous engine behind faster, less power-hungry phones and computers, so if this scales it eventually shows up as a device that does more without melting your battery, but "prototype" is doing a lot of work between here and there.
Source: engadget.com
wired.com●Covered by 3 sources
GTA VI breaks the $70 ceiling, and the disc in the box might not even hold the game
After more than a decade of waiting, Grand Theft Auto VI is finally near release and will reportedly cost more than the standard $70 AAA tag, while the physical edition is drawing fire as DRM in a box rather than a real disc you own.
Source: wired.com
SecurityWeek●Covered by 4 sources
Anthropic's AI found holes in classified US systems, so the government pulled the plug on it
A US official says Anthropic's Mythos model, tested with intelligence agencies under Project Glasswing, quickly turned up vulnerabilities in classified systems, after which the Trump administration slapped on export controls that cut parts of the NSA off from Mythos 5. Three House members (Liccardo, Obernolte, and Lieu) gave Commerce until June 26 to explain when the public regains access, and Legion has filed the first legal challenge arguing hosted model access is not the same as exporting weights.
So what? ↓The tool worked so well that Washington's reflex was to restrict it, which tells you the fight over who's allowed to use frontier AI is now a legal and political mess that decides whether you ever get to touch the good models.
Source: SecurityWeek
OpenAI●Covered by 3 sources
OpenAI's most-used ChatGPT model gets an upgrade, free tier included
OpenAI is rolling out GPT-5.5 Instant inside ChatGPT for both paid and free users, with the company saying the update improves intent recognition, context across multiple turns, and handling of complex multi-condition prompts. By OpenAI's own description, naturally, since that's the only benchmark on offer here.
So what? ↓This is the default model most people actually type into, so if it really follows multi-step requests better, the upgrade lands on the free account in your pocket, not just some enterprise pilot.
Source: OpenAI
NVIDIA●Covered by 3 sources
NVIDIA's NeMo AutoModel claims 3.7x faster MoE fine-tuning and 32% less GPU memory
NVIDIA put NeMo AutoModel on Hugging Face to speed up fine-tuning of Mixture-of-Experts models, citing up to 3.7x higher training throughput (via Expert Parallelism, DeepEP, and TransformerEngine kernels) plus a 32% cut in peak GPU memory, all per NVIDIA's own numbers.
So what? ↓GPU memory and training time are the two bills that decide whether a small team can fine-tune a big model at all, so if these figures hold up outside NVIDIA's benchmarks, more people get to touch the expensive end of AI without renting a data center.
Source: NVIDIA
Interconnects●Covered by 3 sources
GLM-5.2 matches Claude Opus at one-fifth the price, and it's open-weight
Zhipu AI's GLM-5.2 tops open-model rankings on Artificial Analysis and Agent Arena, scored 22.8% on ARC-AGI-2 (the best yet from an open model), and a Snowflake benchmark of 103 coding tasks found it nearly matching Claude Opus 4.7 at a fifth the cost per output token. The catch: it burns roughly twice as many tokens per task, so the gap narrows once you do the math.
So what? ↓An open-weight model trading blows with the priciest closed labs at a fraction of the price is exactly the kind of competition that drags everyone's AI bills downward, including yours.
Source: Interconnects