1.
[AINews] Muse Spark 1.3 matches GPT-5.6-Sol, confirming Meta Superintelligence as the newest Frontier Lab, >90% discount for training
So, Muse Spark 1.3 just dropped, and it’s got everyone buzzing because it’s now ranked as the third-best AI model globally. It’s finally showing some serious competition against the likes of OpenAI and Anthropic, which is pretty exciting. What’s interesting is the pricing model; if you opt into training, you can get it for over 90% off. That’s a big deal for developers looking to experiment without breaking the bank.
On the educational front, Stanford is shaking things up by formalizing AI-native software engineering. They’re revamping their curriculum to focus on building agents from scratch, moving away from just prompting. This shift suggests a deeper dive into systems-oriented engineering, which could really change how we think about AI development.
Meanwhile, there’s chatter about the Astra architecture being a “looped transformer,” but it seems to be more of a modest tweak rather than a huge leap forward. It’s all about optimizing memory and compute without sacrificing too much efficiency. Plus, there’s a push for real-time multimodal workloads, with updates like Photon 2.1 enhancing text-to-speech capabilities.
Lastly, there’s some concern about the effectiveness of skill retrieval in AI tasks. While it might boost overall performance, it can actually hurt specific tasks where those skills are applied. It’s a reminder to be cautious about interpreting aggregate improvements. Overall, it feels like we’re at a fascinating crossroads in AI development, with new models and educational shifts paving the way for a more nuanced understanding of how these systems can work together.
2.
We’re Never Slowing Down Again
For most of history, wealth was a quality dial. A king rode the fastest horse, ate the freshest food and saw the best physician, and each rung down the ladder got a slightly worse version of all three. Money bought a better thing. That is what money was for. For the most remarkable products humans now make, the dial is gone. There is a car that will take you from your driveway to a city you have never seen without your hands on the wheel. It sells for about what a Honda Civic does, and the software that drives it rents for $99 a month.
3.
Anthropic’s Fable 5.1 guide reads like a manual for agent product design, not a collection of prompt tricks
Anthropic just rolled out Claude Fable 5.1, and it’s more than just a faster model; it’s a guide for how to design effective agent products. The key takeaway is that the user experience hinges on how well the model, prompts, and system work together. They emphasize that a model upgrade doesn’t guarantee an improved product unless the entire system supports it.
The guide suggests starting with a high effort setting and adjusting based on task needs, which shifts the focus from just quality to more efficient routing decisions. It also highlights that users might not see visible updates during long tasks, so it’s crucial to adjust client settings to ensure users receive meaningful progress updates.
Moreover, Fable 5.1 can batch calls, but it’s important to prompt the model to identify and request multiple items in one go to avoid unnecessary delays. Lastly, they stress the importance of maintaining a stable conversation prefix to keep reasoning intact. These insights are not just about prompts; they’re fundamental to how the product architecture can shape user experience.
4.
9/2: OpenAI’s Astra Uses Recurrent Depth Technique
OpenAI’s Astra slipped a new “recurrent depth” trick under the hood: it shoves more of the reasoning into hidden activations instead of the visible chain‑of‑thought text. The model doesn’t get bigger, but it nudges performance up, which has folks worrying about monitorability. OpenAI says the reasoning is still traceable and that we’ll need fresh tools to keep an eye on it.
Google rolled out Gemini 3.8 Flash and a “Cyber” variant that relaxes some security limits for its Fairwind partners, keeping the same pricing as the previous flash. Meta’s Muse Spark 1.3 trims hallucinations, cuts tool calls by a fifth and uses fewer tokens, even edging Gemini on a few benchmarks. Alibaba’s Qwen 3.8‑Max‑0902 and Europe’s Quasar 438B both push token windows to a million, with the latter hitting a low price point.
Meanwhile, the DOJ slipped a brief to OpenAI saying training on copyrighted text is likely fair use, the first formal U.S. stance on the issue. Anthropic got a nod from the Trump administration again, and the three big labs—OpenAI, Anthropic, DeepMind—are teaming up on safeguards to let legit biology research happen while blocking virus‑design assistance.
5.
Meta Pushes Its New AI Agent on Employees—but Eases Off on Tokenmaxxing
Meta is formally ending what amounted to a tokenmaxxing incentive program for employees. In an internal announcement this week, the social media giant told workers that their performance evaluations would no longer be dependent upon how much they used AI tools, three employees who received the message tell WIRED. But as Meta workers simultaneously begin testing a new agentic AI tool known as Hatch, they say their token consumption continues to surge.
6.
Microsoft, Famous for Culty Jargon, Is Naming Half Its Organization ‘Agents and Infra’
There’s the already irksome world of general purpose corporate jargon, and then there’s Microsoft, where dedication to saying things in a weird way borders on the religious. Add AI to the mix, and you get today’s news: According to the Wall Street Journal, Microsoft will no longer be divided into these three passé reporting segments: - Productivity and Business Processes - More Personal Computing Now the org chart will have just two hip, young divisions: The Journal notes that according to CEO Satya Nadella, AI in particular is changing the structural boundaries within Microsoft.
7.
Canonical joins the Open Secure AI Alliance
8.
When is Apple releasing new AirPods?
Apple’s AirPods lineup is stronger than ever, but there are already rumors about what’s coming next. Here’s everything we know so far about when Apple will release new AirPods. Apple’s current AirPods lineup Here is Apple’s AirPods lineup as it stands today: - AirPods 4 (currently on sale for $99) - AirPods 4 with Active Noise Cancellation (currently on sale for $149) - AirPods Pro 3 (currently on sale for $99) - AirPods Max 2 (currently on sale for $479) When is Apple releasing new AirPods? The biggest change coming to Apple’s AirPods lineup is a new version of AirPods with built-in cameras.
9.
Dell’s $95B AI backlog shows the infrastructure crunch is far from over
Dell Technologies is acknowledging that infrastructure and storage supply still can’t keep up with agentic AI’s insatiable appetite for resources. The company this week reported a “record” AI backlog, with $95 billion in orders waiting to be filled. This dovetails with quarterly earnings reflecting a more than 50% year-over-year increase in AI demand.
10.
Lemonade 11.9 Local AI Server Released With Super Exciting AMD ROCm HRX Backend
Lemonade 11.9 is out today as the newest feature release to this AMD-backed, open-source local AI server solution across Linux, Windows, and macOS. Lemonade has long been focused on offering "100% free and private" AI use with local hardware whether it be GPUs, CPUs, or NPUs. With Lemonade 11.9's release today it's very interesting for having experimental ROCm HRX back-end support with Llama.cpp. HRX is the new exciting thing to watch out for on the AMD ROCm compute landscape...