Among the increasing concern about screen time in school comes a new culprit: the vetting process for school software.

A growing group of parents and teachers has spent the last few years fighting against cellphones in the classroom, with some extending that to all digital devices. But the school-issued laptops, and the software accompanying them, have been left largely unscathed.

“A lot of the issues with personal devices can move to the district-issued devices,” said Kim Whitman, co-lead for Smartphone Free Childhood US, in a previous interview with EdSurge. Whitman explained that when students do not have cellphones, they can still message with friends on their Chromebooks, or through tools like Google Docs. “There are definitely issues with school-issued devices as well.”

Proposals in three states – Rhode Island, Utah and Vermont – are now tackling these concerns.

Better Vetting Processes

At the start of this year’s legislative session, all three states concurrently proposed reviewing the vetting process of education software.

In most districts, school boards, IT personnel and administrators choose vendors, often relying on the vendors’ own data to prove the products’ safety and efficacy.

“There is nobody right now that is confirming these products are safe, effective and legal,” Whitman said in a previous interview. “It should not fall on the district’s IT director; it would be impossible for them to do it. And the companies should not be tasked with doing it — that would be like nicotine companies vetting their own cigarettes.”

The proposed legislation is looking to change that.

Vermont

Bill: An act relating to educational technology products

Status: Passed by the House March 27; currently before the Senate Committee on Education

This bill proposes to require that providers of educational technology products register annually with the state. It also requires the secretary of state to create a certification standard and review process for these products before schools can use them.

Any provider of an educational technology product — specifically student-facing tools that are used for teaching and learning in schools — must register with the secretary of state, pay a registration fee of $100 and provide its most up-to-date terms and conditions and privacy policy.

The secretary of state would work with the Vermont Agency of Education to review registrations.

Criteria for certification include:

• The product’s compliance with state curriculum standards
• Advantages of using it versus non-digital methods
• Whether it was explicitly designed for educational purposes
• Design features, including artificial intelligence, geotracking and targeted advertising

While the initial bill proposed that any edtech provider not certified by the state, but continues to operate, could be liable for fines of $50 a day up to $10,000, that language was struck by the final bill passed from the House.

If passed by the Senate, the bill would go into effect July 1, 2026. By November 2027, the Agency of Education would submit a written report on which state entities should be involved in the edtech certification and any other recommendations for certification.

Utah

Bill: Software in Education

Status: Signed into law on March 18

The bill requires the Utah Board of Education to study the use of software and digital practices in public schools, review best practices and provide guidance for responsible use.

The state also passed a Classroom Technology Amendments bill tackling screen time at every grade level, banning it entirely from kindergarten through third grade, except for computer science and assessments. Middle school students must have their parents “opt-in” to taking devices home and high school students will be allowed to bring home devices unless parents “opt-out.”

“We’re not anti-technology,” Rep. Ariel Defay (R-UT) said in a statement. She is a sponsor of the Classroom Technology Amendments bill. “We just want to ensure that education technology is used intentionally and actually helps students to learn.”

Rhode Island

Bill: The Safe School Technology Act of 2026

Status: Passed by the House April 14; currently in the Senate Education Committee

This bill, proposed by three Rhode Island representatives who are also mothers, is part of a six-bill package focused on protecting children from social media, artificial intelligence and digital platforms.

The Safe School Technology Act bill would be enacted this August if approved, banning software providers from activating or accessing any audio or video functions on a device outside of school-related activities. It also bans the use of location data.

The initial bill lists a litany of concerns that the “lack of regulation” caused, including increased screen time, and “marketing commercial products as educational with no accountability; children being given devices without proof of developmental appropriateness and parents being excluded from decisions about their child’s digital exposure.”

But the main concern, argued by state Representative June Speakman (D-RI), who sponsored the bill, is that a majority of school districts’ technology policies do not have limits on tracking student devices. She added roughly two-thirds of districts also do not limit school-issued device’s ability to activate audio and video.

“Passing this bill will provide clear, consistent protection across all schools in the state that assures students and their families that their devices cannot be used to invade their privacy or track their activities,” Speakman said in a statement.

“They deserve to feel confident that their privacy is protected when they use technology that is required for school,” she added.

Tech Pushback

Several technology proponents have pushed back.

The Software and Information Industry Association spoke out against the Rhode Island bill in March, saying if the bill passed it would make the state be one of the most restrictive in the nation.

In an open letter to Joseph McNamara, chair of the Rhode Island House Education Committee, Abigail Wilson, director of state policy at the Software and Information Industry Association, said the bill “proposes an overly restrictive regulatory framework that will severely disrupt classroom instruction, impose massive unfunded administrative burdens on local schools, and deprive Rhode Island students of critical, evidence-based learning tools.”

Keith Krueger, CEO of the nonprofit Consortium for School Networking, told NBC News that the proposed legislation “does keep me up at night.”

“I think some well-intentioned policymakers … are rushing so quickly that they haven’t thought through the implications,” he said.

• Panthalassa’s valuation now sits near $1 billion after fresh funding
• Peter Thiel led a $140 million investment round into the ocean tech company
• Investors see ocean energy as a vast, untapped computing resource

A US-based ocean technology company, Panthalassa, is advancing its plan to relocate data processing into open waters, backed by fresh funding that places its valuation near $1 billion.

The start-up has spent ten years developing wave energy technology and is now backed by PayPal co-founder and early Facebook investor Peter Thiel, who led a $140 million investment round into the company.

Kodiak AI’s stock tumbled 37% in after-hours trading Thursday after the self-driving truck startup disclosed it had raised $100 million by selling shares at a steep discount — a sign that investors were willing to back the company but not at its current market price.

The company sold shares at $6.50 each, well below its closing price of $9.10, according to a filing with the Securities and Exchange Commission (SEC). The raise also included warrants — instruments that give investors the right to buy additional shares later at a set price, in this case as low as $6.

The financing came from existing backer Ares Management and several unnamed institutional investors.

The influx of capital comes Kodiak pushes forward on the expensive task of scaling its self-driving trucks business, which covers off-road industrial settings and public highways, with the ultimate goal of eventually spending less than it earns. Kodiak reported revenue of $1.8 million in the first quarter, up from the $1.4 million it logged in same period a year prior. The company’s loss from operations was $37.8 million, twice what it reported in the same period last year.

Those numbers help explain why the discount terms rattled investors. The company is burning cash fast, and the raise — while sizable — does little to change that math in the near term.

Kodiak has made some recent progress on the business front, including a new commercial contract with Roehl Transport, a pilot program to test Kodiak-equipped autonomous trucks at West Fraser Timber Co.’s log-hauling operations in Alberta, Canada, and a collaboration with the military vehicle maker General Dynamics Land Systems to create autonomous ground vehicles for defense applications.

Under the deal with Roehl, which was also announced Thursday, Kodiak-equipped trucks will autonomously haul freight between Dallas and Houston on four round trips per week. The trucks operate autonomously on the entirety of the trip, but Kodiak keeps a human safety operator behind the wheel as a precaution.

Kodiak founder and CEO Don Burnette said the company is on track to move to driverless trucking on public highways later this year as it ramps up operations.

“We have tons of over-the-road long haul initiatives, and bringing on new partners continues to show momentum,” he said in an interview. “We’re excited about the progress that we’re making as we march toward our driverless launch later this year.”

For now, Kodiak owns the trucks, provides the safety driver, and carries the freight for Roehl along with its other existing on-highway customers, which include Werner, J.B. Hunt, Bridgestone, Martin Brower, and C.R. England. But that arrangement will change once it goes to driverless trucking operations.

“Our intention is to not own the trucks at that point [but to] operate our driver-as-a-service model, where [customers] own and operate the trucks,” Burnette said. He added that this is the system it uses with its off-highway customer Atlas for its driverless deployment in the Permian Basin of Texas.

While Kodiak plans to pull the safety driver by the end of 2026, Burnette said it won’t start driverless operations on public highways until it has finished validating the technology.

“It’s already operating under all of the conditions that we expect to launch driverless, but there’s a lot of validation work that we need to do, and that’s where we bring in our autonomy readiness measure,” Burnette said, describing the initiative — released Thursday — as a zero-to-100 score tracking how much of Kodiak’s internal safety validation is complete. As of April, Kodiak was at 86%, Burnette said.

The company, which was previously called Kodiak Robotics, went public in September via a merger with special-purpose acquisition company Ares Acquisition Corporation II, an affiliate of Ares Management. The deal valued the startup at about $2.5 billion. 

At the time, Kodiak raised $275 million in financing. More than $212.5 million came from certain institutional investors, including $145 million in PIPE funding (Private Investment in Public Equity, a method by which investors purchase shares directly from a public company) and about $62.9 million in trust cash from Ares. That trust cash shrank from its initial $562 million as some SPAC investors redeemed their shares — a standard provision that lets SPAC investors recover their money before a merger closes.

Samsung has started rolling out One UI 8.5, and it’s a pretty big one.

The update is landing on more than 40 Galaxy smartphones and tablets. It brings a refreshed interface and a new wave of AI features to devices that, in some cases, are years old.

The rollout began in Korea today, with other regions expected to follow from mid-May. As usual with Samsung updates, it won’t arrive everywhere at once. Even though the list of supported devices is already extensive, some users will still have to wait.

At the front of the queue are Samsung’s newest flagships, including the Galaxy S25, S25+, and S25 Ultra, along with the Galaxy S25 FE and S25 Edge. But the update doesn’t stop there; Samsung is also pushing One UI 8.5 to older generations like the Galaxy S24 and S23 series. This includes their Ultra and FE models.

SQUIRREL_PLAYLIST_10206388

Foldables are included too, with support for the Galaxy Z Fold 5 through Fold 7 and Galaxy Z Flip 5 through Flip 7. Additionally, a wide range of tablets including the Galaxy Tab S9, S10, and S11 series will be supported. Even mid-range and entry-level devices like the Galaxy A05 are on the list.

That breadth is the main story here. Rather than limiting new software to premium devices, Samsung is once again pushing its latest One UI version across almost its entire ecosystem.

So what’s actually new? One UI 8.5 brings a visual refresh in parts of the interface, including updated menus and navigation elements. But the bigger focus is AI. Samsung is expanding tools like Photo Assist, which helps refine and adjust AI-generated images, alongside improvements to its Bixby assistant, which is becoming more context-aware and responsive.

It’s not a radical redesign, but it does continue Samsung’s steady shift toward AI-driven features across both hardware tiers and older devices.

As with most major Android updates, availability will vary depending on region and model. It may take months before every eligible Galaxy device receives it. Still, for a rollout that includes everything from flagship phones to budget A-series models, this is one of Samsung’s broader software updates in recent memory. In effect, it is a free upgrade sitting in the settings menu for millions of users.

Project Hail Mary introduced fans to an unforgettable alien named Rocky. Many who finished the book wanted more time with the character and his quirky way of speaking. One maker decided to satisfy that craving by constructing a physical robot that captures the essence of Rocky in every joint and word.

The engineer behind Leviathan Engineering worked tirelessly on this project for months, bringing Rocky to life. He began with digital models of the character purchased from 3D Totems, a store that does an excellent job of creating 3D models that are precisely perfect. Next, software programs such as Fusion 360 and Tinkercad were used to ensure that the parts were not only printable, but also strong enough to endure some hard treatment.

Months later, the printed components came together to create the body of a four-legged monster with arms that appear to lunge out at you in all the right ways. Ten metal-geared servos provide the driving force behind the movements, cleverly placed to allow the robot’s expressiveness to show. The shoulders each receive an additional servo for arm swings, while the knees of the legs receive one each for low crouches. Movement, gestures, and body language pull the alien’s exuberant personality straight out of the novel. He even gets to offer you a full fist bump, or make a wild arm gesture, just like in the novel.

The robot is powered by a Raspberry Pi 5 that is connected to a circuit board known as the PCA9685 HAT, which operates the servos and controls all of their motions. Power comes from an external supply because the motors draw plenty of current during lively movements. Software brings everything to life. It has speech recognition integrated with Vosk, so you can tell it what to do without experiencing internet lag. Then there’s Piper for the robot’s voice, which has that really unique, staccato talking style that we all love about Rocky. For talks, however, it uses Google’s Gemini model, which essentially determines what the robot says and even what gestures to make based on the situation.

The maker wrote all the code using assistance from Claude through its command line interface. No fixed animation scripts exist. Instead the language model chooses movements based on context through a process called tool calling. Ask for a fist bump and the arm extends while the robot says something like “fist bump yes much happy.”

Of course, like any good maker project, this one had some challenges along the way. The engineer experimented with pulleys and linear actuators before settling on servos since they provided him more control over the entire mess. To ensure that printed joints did not break under any force, he had to do some trial and error before getting it perfect. Then there was the enjoyable task of putting things together, a little hot glue here and little super glue there to keep everything from falling apart. Wires route neatly inside the body thanks to extension cables. The final assembly stands about the size of a small tabletop model yet moves with enough grace to feel like a living creature from the pages of the novel.
[Source]

If you talk to the FDA, there’s only one permanent method of hair removal—electrolysis. This involves sticking a needle into a hair follicle, getting it very hot or running a current through it, and then letting heat and/or the lye generated kill the root of the hair dead. Normally, you’d pay someone with a commercial machine to do this for you at great expense. Or, you could do it yourself with a home-built machine, as [n3tcat] did.

Based on the available information out in the wild, [n3tcat] decided to build a galvanic electrolysis machine. This specifically passes current through a needle in the hair follicle to generate lye at the hair bulb, which kills it. The amount of lye generated depends on the amount of current and the time over which it is applied. More lye is more likely to kill a follicle permanently, though there are limits with regards to avoiding scarring, other skin damage, and excessive pain.

[n3tcat]’s guide explains the basic theory behind galvanic electrolysis, as well as how the rig was built. An early attempt simply involved hooking up a 12-volt car battery to a standard electrolysis needle, sticking it in a hair, with the other electrode being an aluminium can held by the person being treated. The fun thing was that this allowed varying the current depending on how much contact and how stiffly the person grabbed the can.

After a few successful hair removals this way, [n3tcat] decided to build a better rig. An RP2040 microcontroller was enlisted to run the show, powered by a 3.7-volt lithium rechargeable battery. An OLED screen and a rotary encoder were selected to serve as the interface, while a foot pedal was added for firing off current. A boost converter was used to push the battery voltage up to the vicinity of 15 volts for delivery to the needle, set up to avoid excessive current delivery for safety. A DAC was paired with an LM358 op-amp feeding into a MOSFET to control the current passed to the needle for accurate, controlled treatment, with the RP2040 monitoring the current level via a dedicated ADC. The needle itself got a D-printed pen-like handle for better ergonomics, easing the process of slotting the needle into a hair follicle. Everything was then assembled on a cute PCB, and wrapped up in a nice 3D printed housing. The files are available for the curious.

Electrolysis is a process that can cost many thousands of dollars depending on how much hair you hope to remove. Thus, it’s easy to see the appeal in having a rig that lets you do it at home. It’s just one of those things where you have to take the proper precautions to ensure you’re not unduly hurting yourself.  Stay safe out there, hackers!

Even as leading AI providers like OpenAI and Anthropic battle over the compute to train and release ever larger, more powerful models, other labs are going in a different direction — pursuing the development of smaller, more efficient models and often open sourcing them.

The latest worth paying attention to comes from the lesser-known Palo Alto startup Zyphra, which this week released its new reasoning, mixture-of-experts (MoE) language model, ZAYA1-8B, with just over 8 billion parameters and only 760 million active — far fewer than the trillions estimated for the likes of the big labs. Yet, ZAYA1-8B retains competitive performance on third-party benchmarks against GPT-5-High and DeepSeek-V3.2.

It can be downloaded from Hugging Face now free of charge under a permissive, standard, enterprise-friendly Apache 2.0 license — and enterprises and indie developers can begin using and customizing it immediately to suit their needs. Individual users can also test it themselves here free at Zyphra Cloud, the startup’s inference solution.

But the real headline is what ZAYA1-8B was trained on: a full stack of AMD Instinct MI300 graphics processing units (GPUs), the rival to Nvidia GPUs released by AMD nearly three years ago, and which shows that this platform is capable of producing useful models and is a viable alternative to the preferential position Nvidia has maintained in recent years among AI model developers.

How ZAYA1-8B was trained

The “intelligence density” touted by Zyphra is the result of what they describe as a “full-stack innovation” approach, spanning architecture, pretraining, and reinforcement learning (RL).

ZAYA1-8B is built on Zyphra’s proprietary MoE++ architecture, described in a technical report released by the lab. This architecture introduces three fundamental changes to the standard Transformer architecture that gave rise to large language models (LLMs) and the entire generative AI era:

1. Compressed Convolutional Attention (CCA): Unlike standard attention mechanisms that struggle with memory as context windows grow, CCA performs sequence mixing in a compressed latent space. This results in an 8x reduction in KV-cache size compared to full multi-head attention, enabling more efficient long-context reasoning.

2. The ZAYA1 MLP Router: Most MoE models use a linear router to decide which “experts” handle a specific token. Zyphra replaced this with a more expressive multi-layer MLP-based design. To maintain stability during training—a common hurdle for MoEs—they implemented a bias-balancing scheme inspired by PID controllers from classical control theory.

3. Learned Residual Scaling: This controls the growth of the “residual norm” as data flows deeper into the model’s 40 layers, preventing gradient vanishing or explosion with negligible computational overhead.

Reasoning-First Pretraining

A critical differentiator for ZAYA1-8B is that reasoning was integrated from the start of pretraining, rather than being “bolted on” during post-training.

To handle long chain-of-thought (CoT) traces that would otherwise exceed the initial 4K pretraining context, Zyphra developed Answer-Preserving (AP) Trimming.

Think of AP-trimming like a film editor cutting a long scene: instead of cutting the ending (the solution) or dropping the scene entirely, the editor removes the “middle” of the character’s monologue while keeping the beginning (the problem setup) and the final reveal (the answer).

This ensures the model learns the relationship between complex problems and their solutions even when the full internal logic doesn’t yet fit into memory.

It seemed to work well on my test query about countertop stain removal to ZAYA1-8B running on Zyphra Cloud.

Markovian RSA: redefining test-time compute

The model’s most significant performance leap comes from Markovian RSA, a novel test-time compute (TTC) methodology.

Traditionally, if you want a model to “think harder,” you let it generate a longer chain of thought. However, this often leads to “context bloat,” where the model loses focus as the history grows too long.

Markovian RSA solves this by decoupling “thinking depth” from “context size”. It functions like a recursive scientific peer-review process:

• The model generates multiple parallel reasoning traces (candidates).

• It then extracts only the “tails” (the last few thousand tokens) of these traces.

• These tails are subsampled and presented to the model in a new “aggregation prompt,” asking it to reconcile the different approaches into a better solution.

By carrying forward only the tails (typically a 4K-token budget), the model can reason indefinitely without the context window ever overflowing. In practice, this allows the 700M active parameter ZAYA1-8B to achieve a 91.9% score on AIME ’25, closing the gap with models that have 30 to 50 times its active parameter count.

Because ZAYA1-8B maintains a small total parameter footprint (8.4B), it is uniquely positioned for on-device deployment and local LLM applications. For enterprises, this enables the deployment of high-tier reasoning capabilities—traditionally reserved for massive cloud-based models—directly onto local hardware or edge devices. This “local-first” reasoning approach addresses common enterprise hurdles regarding data residency, latency, and the high cost of persistent API dependencies.

Benchmarks show a remarkably performant small model that punches above its weight class

Zyphra is positioning ZAYA1-8B as a “punch above its weight” model for developers who need high-tier reasoning without the latency or cost of massive frontier models. After all, its active parameter count is much lower than other similarly-sized models, making it much cheaper and less compute-intensive to run in inference.

• Instruction Following: ZAYA1-8B scores 85.58 on IFEval, remaining competitive with much larger models like Intellect-3 (106B).

• Agentic Capabilities: On the τ² benchmark, the model reaches 43.12, and 39.22 on BFCL-v4, providing a baseline for its ability to handle tool-calling and multi-turn tasks.

In single-rollout evaluations (without the extra “thinking” time), ZAYA1-8B already outperforms its weight class. It beats Qwen3.5-4B and Gemma-4-E4B on math and code benchmarks.

When Markovian RSA is enabled, the results are startling:

• HMMT ’25 (Math): ZAYA1-8B hits 89.6%, surpassing Claude 4.5 Sonnet (79.2%) and GPT-5-High (88.3%).

• LiveCodeBench (Coding): The model achieves 69.2%, outperforming DeepSeek-R1-0528.

Zyphra notes that while the model is a specialist in algorithmic reasoning, it lags slightly behind larger models on “knowledge-heavy” tasks like broad factual retrieval (MMLU-Pro), which suggests that while reasoning can be compressed into smaller cores, factual memory still benefits from raw parameter count.

Apache 2.0 open licensed for research and commercial usage

Zyphra has released ZAYA1-8B under the Apache-2.0 license. This is a critical choice for the developer community. Unlike “copyleft” licenses like the GPL, which require any derived work to also be open-source, Apache-2.0 is highly permissive.

For developers and enterprises, this means they can use, modify, and distribute ZAYA1-8B—even within proprietary, commercial applications—without being forced to open-source their own codebases.

It also includes an explicit grant of patent rights from contributors, providing a layer of legal safety for startups building on top of Zyphra’s architecture. By opting for Apache-2.0 over more restrictive “research-only” licenses often seen from frontier labs, Zyphra is signaling a commitment to the open-weight ecosystem.

To deploy ZAYA1-8B, developers must use specific branches from Zyphra’s forks of core libraries, as the architecture requires specialized handling:

• Custom Forks: Users should install the zaya1 branch from Zyphra’s versions of the vllm and transformers libraries.

• Deployment Flags: When starting a vLLM server, specific flags are required to handle the reasoning parser and tool-calling (e.g., --reasoning-parser qwen3 and --tool-call-parser zaya_xml).

• Parallelism Strategy: For multi-GPU environments, Zyphra recommends using Data Parallelism (DP) combined with Expert Parallelism (EP). Notably, Tensor Parallelism (TP) for the model’s CCA mechanism is not currently supported, making DP+EP the optimal path for scaling inference throughput.

Background on Zyphra

Zyphra: A New Paradigm for Intelligence Density

Founded in 2021 and headquartered in Palo Alto, California, Zyphra Technologies is a full-stack artificial intelligence laboratory dedicated to building human-aligned artificial general intelligence (AGI) — that which outperforms people at most tasks — through a decentralized, open-source framework.

According to the company’s official mission statement, Zyphra seeks to challenge the “centralized” dominance of monolithic cloud models by focusing on “intelligence density”—a core guiding principle that aims to maximize the reasoning and logic extracted per parameter and per FLOP.

Zyphra CEO and Co-Founder Krithik Puthalath explained previously to VentureBeat that this strategy is essential for enabling high-performance AI to run locally on hardware such as tablets, wearable glasses, and enterprise servers, thereby ensuring user privacy and reducing reliance on third-party cloud infrastructure.

The company’s technical identity is deeply informed by computational neuroscience, led by Co-Founder and Chief Scientist Beren Millidge.

According to Millidge’s personal website, he currently serves as a Postdoctoral Researcher at the University of Oxford’s Nuffield Department of Clinical Neurosciences, where his research focuses on deep credit assignment and mathematical models of the brain.

Millidge, who earned his PhD from the University of Edinburgh, has pioneered research into active inference and the “free-energy principle,” concepts that directly influence Zyphra’s pursuit of multimodal architectures capable of long-term memory and continual learning.

This neuroscientific influence was central to the design of Zyphra’s prior Zamba model, released in 2024, which mimics the cortex-hippocampus interaction to share information across sequential layers. A recent TED Talk video provides insight into Millidge’s perspective on the intersection of biological neuroscience and AI, which serves as the theoretical foundation for Zyphra’s model architectures.

Zyphra has achieved significant technical milestones through a deep integration with the AMD hardware ecosystem, as detailed in the company’s research documentation.

Financial data from PitchBook indicates that Zyphra is currently a venture-backed company that attained “Unicorn” status in June 2025 following a $110 million Series A funding round. According to PitchBook and company press releases, Zyphra is supported by a group of strategic investors including Advanced Micro Devices (AMD), IBM, Bison Ventures, and BC VC. With a team of approximately 31 employees as of 2026, the company continues to expand its footprint through the Zyphra Inference Cloud and Maia, an intelligent assistant platform designed to bring advanced search and productivity tools to enterprise teams.

Community reactions and industry context

The announcement has resonated strongly within the AI community, garnering nearly 1 million views on X/Twitter within 24 hours. The excitement largely centers on two factors: the viability of the AMD stack and the efficiency of the reasoning “cascade.”

Technologists have noted that Zyphra’s post-training process—a 4-stage RL cascade—is unusually disciplined. Most labs use a single round of RL, but Zyphra’s pipeline includes a “reasoning warmup” followed by a curriculum of 400 adaptive puzzle-like environments (RLVE-Gym) before finally moving to behavioral polishing.

One of the most praised “under-the-hood” details is Router Replay. In MoE models, training can become unstable if the “trainer” engine and the “inference” engine make slightly different decisions about which expert to use for a token due to floating-point noise. Zyphra’s system records the exact expert choices made during generation and forces the trainer to use them, effectively “pinning” the computation path and ensuring higher learning stability.

As the industry faces a potential plateau in the benefits of simply adding more parameters, ZAYA1-8B provides a compelling counter-narrative: that the next frontier of AI isn’t just about bigger clusters, but about smarter “thinking” algorithms that can do more with less.