• Traction: Beam
  • How To Launch a Laptop
  • How Not to Launch a Laptop

Traction: Beam
Reflection AI is an artificial intelligence startup that was created by two escapees from Google DeepMind. Nvidia is a key financial backer. This week it launched Beam, an open-weight model with more than 500B parameters and a 1 million token context window. Beam is intended to provide assistance with coding and agentic tasks and for that reason seems targeted toward enterprise and business customers. Reflection says that Beam's performance is comparable to popular Chinese open-weight model GLM 5.2 from Z.ai. It is said to outperform other open models while being less demanding computationally for inference, which should lessen the expense.

As an open-weight model with a lower operating cost, Beam could present challenges for American frontier AI labs like OpenAI and Anthropic. While OpenAI and Anthropic are widely considered to have some of the world's most powerful models, the cost to use them is significantly higher than using open weight models and many companies are shifting their focus toward those models not only to save money but to also customize them based on their needs.

As open weight models not only allow for more customization but usually are run locally rather than in the cloud, they can provide greater data safety and security for companies eager to keep private data off external servers. When you combine this with the customization potential and the lower costs, the race to IPO before the market faces radical shifts makes more sense. OpenAI has pushed their IPO back to 2027, but Anthropic is still moving forward with their planned IPO. These companies know that the popularization of open models by users in the United States will continue to put pressure on their businesses and they will need to continue to innovate and devise new methods to increase revenue. This is one of the reasons why all the big AI companies have been pushing agents and agentic platforms. Getting companies and regular people to think they need agents is another avenue to extract recurring subscription fees that are a critical income path and for OpenAI it is an open door to more lucrative enterprise income that has been shoveled primarily in Anthropic's direction up until now.

How to Launch a Laptop
It doesn't seem to have gathered all that much attention this week, but Microsoft held an event this week to reveal their latest Surface, the Surface Laptop Ultra. The laptop features the RTX Spart platform from Nvidia that offers up to a 20-core Grace CPU and up to 6,144 CUDA cores with Blackwell graphics. This is Nvidia's latest and most powerful architecture that they launched earlier this year. The laptop can be configured with up to 128GB of RAM and comes with more ports than any Surface has had in the past - two USB4/USB C, a standard USB 3.1/USB A, a full sized HDMI, an SD card reader, a 3.5mm headphone jack, and a new to-be-patented magnetic USB4/USB C charging port that replaces the old magnetic Surface charging port. The 3:2 touchscreen is 15" with 3270 x 2180 resolution at 262 PPI with 1,000 nits at SDR and up to 2,000 nits at peak HDR. Despite being a laptop, it will support Microsoft's Surface pen. The computer is aimed at local AI processing with the Nvidia platform supporting up to 1 petaflop of 4-bit floating point (FP4) AI performance. Microsoft also launched a Surface RTX box intended for developers. These systems will support local models such as DeepSeek v4 Flash, Nvidia Nemotron, and Microsoft's own MAI Code 1.1 Flash, which is a 137B parameter model that is compressed to reduce memory footprint by ~80%. Microsoft is also bringing its MEXC (Microsoft Execution Containers) to market for Windows 11. MEXC will enable sandboxing features for agentic processes and will provide rails for access to local file systems and controls to shepherd popular agentic systems such as OpenAI Codex, OpenClaw, and GitHub CoPilot.

The laptop features Windows Hello facial recognition, a 92wH battery with up to 15 hours of battery life, a weight of 4.41 pounds, a glass haptic touchpad, and backlit full mechanical keys. A 140W charger is provided in the box. It can be configured with RAM from 24GB up to 126GB and SSD from 512GB to 2TB. Pricing starts at $2,599 for the 24GB RAM 512GB version with the 18-core CPU and 5,120 core GPU. Pre-orders are open now and the machines are expected to ship from October 16, 2026.

At a time when Apple and Linux are vying to be the winner of the platform war for agentic computing, this launch is important for Microsoft to make some noise and provide a platform with their own hardware that developers and AI enthusiasts can use to run powerful models locally. Microsoft supports their hardware well with Surface-specific software updates and their machines are generally known for quality and longevity. They need these new Surface machines to stay competitive in the hardware space. Whether or not they can help them snatch back market share remains to be seen.

How Not to Launch a Laptop
This week has been a rough week for Android fans. After a little bit of fanfare from Google, people were finally able to get their hands on the 5 models of Googlebook available at the initial launch. Most reviewers and influencers on YouTube have been experimenting with the traditional laptop form factor devices from ASUS, Dell and Lenovo. Some have shown off the Acer model, but the HP model has been fairly rare to see. The Dell and HP Googlebooks both have ARM processors while the remaining three devices have Intel architecture powering them. This basic difference presents different experiences right away. Android is optimized for ARM and while it has supported x86 architecture for many years, apps aren't really developed with x86 in mind. Anyone who has paid very close attention to Android over the years was familiar with this difference, but because there have been so few Intel Android devices, the differences really haven't caused any widespread issues over the years. One might have even believed that after all this time the issues were largely resolved. Nevertheless, one would have been mistaken; apparently the issues do remain yet Google decided not to reveal this fact until after pre-orders had shipped and people had gone to stores and bought devices in person and were beginning to use them. Almost as an evil one more thing, Google dropped the news that some apps weren't compatible with the Intel chips that are in 60% of the devices available at launch. But in usual Google fashion, they reduced this to facts and figures: a mere ~300 apps aren't compatible with the Intel Googlebooks. From 9to5Google:

In an email to 9to5Google, Google further clarifies that “96.6% of the top 10k+ most-used applications” in Google Play work “flawlessly” on Intel-powered Googlebooks.

This is where that ~300 figure comes from. But this is only from an unknown list of the most used applications, so whether or not the app you want to run will in fact run is something that will have to be tested app by app. Google says to verify your apps will work before buying a machine - now that people have already bought them - but it's not clear how anyone can accomplish this without just buying a machine and trying things out. No one is going to test 300+ apps just for you. This was something Google should have communicated clearly and honestly upfront. You present the case clearly to users what the benefits and potential pitfalls are for both ARM and Intel processors and let users decide what is best for them. You don't wait until after they've paid their money and taken delivery and perhaps already found that some apps don't work to do a "oh, by the way." It's unkind and unprofessional and Google should have done better.

But the Intel app issue isn't the only issues users have faced this week. From boot looping machines to confusion over the ability to install Linux apps, to weird hardware and software bugs, there have been many things that make it look like the operating system and devices just hadn't been tested enough. My own first day with my Acer Googlebook 14 was filled with apps that wouldn't install, or would install but I couldn't sign in, or apps I could install and sign in but then they'd immediately crash, buttons I couldn't push and advertised features that didn't work, like connecting my Samsung Galaxy S25 Ultra to take advantage of screen mirroring and continuation features. Google did another "oh, by the way" with that, too; apparently Samsung devices are not yet compatible with any of the sharing features and anyone wanting to use those features now will need a recent Pixel running Android 17. Fortunately, I have a Pixel I can use to test, but for anyone who very reasonably owns only one phone, if it's not the right one, then you have bought into a system whose full value you can't even tap into yet. Once again, Google owed it to consumers to tell them honestly and clearly what features would and would not be available at launch and even if this issue is 100% Samsung's fault (it's not clear what the issue really is), partner with them to communicate as best as you can about how long it will take to get the updates people will need to make full use of a Googlebook with their Samsung phone. Samsung is the most popular OEM for Android phones in the United States. You can't just pretend these issues don't exist and hope people won't notice.

Rumors are swirling on the internet that Samsung is working on its own Android laptop and that is why they have sat out Googlebook for now. Perhaps it was a wise move.