In partnership with
Welcome, AI enthusiasts
American startup Reflection is entering the race with Beam, a new open model built to compete with GLM and Kimi. Its downloadable weights arrive this month, giving teams a U.S. option they can run and customize themselves. For companies weighing cost against who supplies their AI, that’s a choice worth watching. Let's dive in!
In today’s insights:
New U.S. Open Model Challenges China’s GLM and Kimi
OpenAI Will Add Invisible Text Watermarks in the EU
Higgsfield’s New Ad Studio Makes Ads for Cents
Read time: 4 minutes
LATEST DEVELOPMENTS
MODEL LAUNCH
🧠 New U.S. Open Model Challenges China’s GLM and Kimi
Evolving AI: Nvidia-backed startup Reflection has unveiled Beam, a coding and reasoning model built to compete with China’s leading open AI systems.
Key Points:
Reflection reports reasoning performance comparable to GLM-5.2 with roughly one-third to one-quarter of the estimated computing work.
Beam activates just 23 billion of its 501 billion parameters, the settings learned during training, for each piece of text it generates.
Downloadable weights are due this month under Apache 2.0, allowing developers to run, modify and use Beam commercially.
Details:
Reflection was founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. The team trained Beam from scratch on 23.8 trillion tokens, then used 10,500 Nvidia chips for four weeks of reinforcement learning, rewarding successful solutions while discouraging unnecessary reasoning. The company is targeting businesses and governments that want a capable U.S. alternative to Chinese open models they can run and adapt themselves.
Why It Matters:
Reflection’s Beam could weaken the case for paying a premium for closed U.S. models. Chinese open models have already won American customers with lower costs and flexibility, despite political scrutiny of their suppliers. A capable U.S. alternative would give businesses and government agencies more choice while preserving the freedom to customize their AI. With another model already in training, Reflection could pressure competitors on both sides to offer better value, giving teams more options when their AI bills outgrow their budgets.
TOGETHER WITH GUIDELESS
🤖 Your software changed. Your training shouldn’t start from scratch.
Evolving AI: Traditional training content becomes outdated fast. A new button, updated workflow, or changed interface often means re-recording videos, replacing screenshots, and rewriting instructions.
That’s where Guideless is different. Instead of creating static training content, Guideless turns your software workflows into guides that are easy to manage, update, and reuse as things change.
With Guideless, you can:
Update individual steps instead of rebuilding an entire guide
Edit instructions, screenshots, and narration in one place
Create natural-sounding voiceovers in 40+ languages
Share or embed guides wherever your team or customers need them
Manage all your software training from one centralized platform
So when your software changes, your training can change with it without starting over. And because Guideless does up to 90% of the creation work for you, getting started doesn’t require a big documentation project either.
Create an account for free, capture your first workflow, and see how quickly Guideless turns it into a finished guide.
Use code EVOLVINGAI to get 50% off any plan on your first purchase.
Source: The Decoder
Evolving AI: OpenAI will add invisible watermarks to eligible ChatGPT and Codex responses in the EU, helping identify text its models generated or edited.
Key Points:
Eligible EU users across all plans will receive watermarked text over the coming weeks, while developers worldwide can already opt in for select models.
In one OpenAI test, replacing 25% of words with synonyms cut watermark detection from about 92% to 17%.
The text detector will initially be available only to approved researchers and expert organizations, with no public access at launch.
Details:
OpenAI is rolling out textGrain to comply with EU rules requiring AI-generated text to be identifiable by software. The system creates a statistical pattern through word choices without adding hidden characters or visible marks, and internal tests found no meaningful impact on quality or speed. A watermark can suggest that OpenAI generated or edited a passage without showing how much a person contributed, while a missing watermark leaves open the possibility that AI was involved.
Why It Matters:
OpenAI’s rollout shows how EU rules are changing everyday writing tools, while the company also controls who can verify the marks it adds. That makes independent testing important before schools or employers rely on the results. People may use AI to express their own ideas more clearly or communicate in a second language, and a watermark cannot explain that contribution. If these checks eventually influence access to education or employment, people will need a fair way to challenge mistakes before a judgment about AI use limits their opportunities.
Evolving AI: Higgsfield’s Ads Studio uses your website to write and design image ads, giving businesses several creative directions to test.
Key Points:
Businesses can generate up to 20 ad variations for each product and select up to 10 products for a single batch.
Each ad costs 4 credits, roughly $0.20 using Higgsfield’s published estimate, putting a batch of 20 ads at about $4.
A saved brand kit stores your logo, colors, fonts, and tone of voice, so future ads use the same settings without repeating the setup.
Details:
Higgsfield already offers image and video ad tools, and Ads Studio adds a workflow built around different reasons to buy a product. For a coffee brand, one ad might highlight the ingredients and another the morning routine, with the copy and visuals changing together. Users can request new directions through chat, including through Higgsfield’s ChatGPT extension, then review and download their choices to run on an advertising platform.
Why It Matters:
Higgsfield gives small businesses a cheap way to explore different sales pitches and more control over routine promotions. An owner can request another version without commissioning a fresh round of copy and design work. That could put pressure on agencies charging mainly for basic ad production, making their understanding of customers and campaign strategy a stronger reason to hire them. For someone running a small shop, producing ads at this cost could leave more of the budget available to reach customers and more room to try an idea they previously couldn’t afford to pursue.
👀 Click on the image you think is real
QUICK HITS
🖼️ OpenAI will test visual ads inside ChatGPT during image generation in the U.S. later this month.
🐛 Google temporarily paused its open-source bug bounty after a surge of automated submissions, most of which it says were invalid.
☁️ AWS added GLM 5.3 to Bedrock, giving developers managed access to the 753B-parameter coding and agent model.
🇰🇷 South Korea plans a $3.5B frontier AI program starting in March 2027 to build a homegrown model that can compete globally.
📈 Trending AI Tools
🎞️ Guideless - AI tool that automatically turns your clicks and workflows into step-by-step video guides*.
🧠 Earlyn - Gives your Mac a private searchable memory of your work.
🎸 FlexChords - Turns any YouTube song into synced guitar chords.
💻 Aperture - Open-source AI code editor that checks changes before applying them.
*partner link