Our report, “The Top 10 Emerging Technologies In 2025,” drops April 29, and this year’s list features big changes. Two longtime list members are stepping back, and two fast-moving entrants will be moving onto the list — including one surprise that wasn’t even on our radar last year until we did the research. These shifts reflect more than changing buzz. They show how client energy, vendor activity, and what we call the AI effect — the rapid spread of AI into everything, everywhere — are reshaping the technology landscape. By publishing earlier than usual this year, we’re giving tech leaders time to absorb the insights, explore strategic implications with us over the summer, and benchmark against the flurry of fall announcements and events. What’s Changing — And Why It Matters Forrester’s annual emerging technologies report identifies the technologies with the most potential to deliver business value in short-, mid-, and long-term horizons. In 2025, we see a clear shift toward AI-adjacent innovation. That’s why we’ve moved two technologies from the top 10 to the next 10: Extended reality (XR) steps aside after years in the top 10. The collapse of metaverse hype and slow progress in consumer-friendly form factors have cooled enterprise interest — for now. But XR still holds promise, and a major breakthrough could bring it back. Zero Trust edge (ZTE) also shifts to the next 10. To reflect market alignment, we’ve adopted the more familiar Secure Access Service Edge (SASE) label. The technology remains important — particularly for security — but it’s no longer one of the top 10. Taking their place will be two fast-rising contenders. The first was once a niche technique but is now vital for training and fine-tuning AI models where real data is scarce or sensitive. It’s powerful — and controversial. Expect more from us soon on its use, risks, and policy implications. The other brings images of science fiction — but thanks to breakthroughs in language models, computer vision, sensing, and declining hardware costs, it’s fast approaching real-world use in manufacturing, logistics, and service jobs. A Comprehensive Emerging Technology Research Experience We’ve expanded well beyond the report. Our emerging technology research portfolio now helps Forrester Decisions clients engage in deeper, more flexible ways: An interactive top 20 experience: Explore each technology’s maturity, impact, and value in an intuitive digital format. In-depth coverage of our top 20: We publish “State Of” and some “Future Of” reports for our top emerging technologies. In addition, we go even further in some cases with “Architect’s Guides” to satisfy our technology architects’ and developers’ desire for detail. Izola, our genAI assistant: Ask questions, extract summaries, or explore implications directly within the research using Izola, now embedded in our emerging tech content. Strategic guidance covering the top 75 technologies: I regularly deliver custom guidance sessions that go beyond the top 10 — and offer emerging tech briefings to executive teams across industries. Don’t Miss The Debut This year’s launch gives you a head start on 2026 strategy. Here’s how to engage: April 29: Forrester clients get access to “The Top 10 Emerging Technologies In 2025” report. May 21: Join our public webinar to hear my first formal presentation of this year’s list. June 23–26: I’ll present deep-dive sessions at CX Summit North America on how emerging technologies are dissolving the barriers between companies and their customers. Curious about what will reshape your business in the decade ahead? This is where to start — and we’re here to help you turn signal into strategy. source

LastPass’ history of data breaches makes it hard to recommend as a viable password manager in 2025. However, for those who wish to try LastPass, this article shows you how to set it up and maximize its available features. How to set up and use LastPass 1. Choose a LastPass subscription LastPass has two categories for its subscriptions: Single Users & Families and Business. LastPass Single Users & Families plans. Image: LastPass For the Single Users & Families plans, you can select between Free, Premium, and Families. If you only need a basic password manager, LastPass Free allows for unlimited password storage and comes with a password generator. However, with the free version, you miss out on important features such as advanced multifactor options, emergency access, and unlimited device access that are only offered with the Premium plan. If you’re a group or a family that wants a centralized password management solution, LastPass Families may be better for you as it covers up to six users. LastPass Business. Image: LastPass Businesses can opt for either LastPass Teams or LastPass Business. Teams covers up to 50 users, while Business is tailored toward larger organizations. These two plans provide more business-centered features, such as shared folders and an administrative console. Fortunately, LastPass offers generous free trials for both categories. In particular, LastPass Premium and Families has a 30-day trial, while LastPass Teams and Business has a 14-day trial. I recommend going for one of these free trials, depending on your needs. The good news is that LastPass doesn’t require users to provide any payment information to access any of these free trials. For this article, I used LastPass Premium’s 30-day free trial. To access one of the free trials, simply click the “Try” button below your subscription of choice. 2. Set up the web app and the browser extension Upon selecting a plan, you’ll be redirected to create your LastPass account. Here, LastPass will ask you to provide an email address and your Master Password. Creating a LastPass account. Image: LastPass In theory, your master password is going to be the only password you’ll have to create on your own. It’s also arguably your most important password, as it serves as the key that unlocks your vault, which contains all your other passwords and credentials. Thus, it’s crucial that you remember your master password. After you’ve finished creating your account, LastPass will ask you to install its browser extension. This is the main way you’ll interact with the service. I use Google Chrome, so LastPass redirected me to the Chrome Web Store. LastPass Chrome browser extension. Image: Google Once you’ve installed the browser extension, you can access LastPass’ web vault application. Before you do this, I recommend you pin the LastPass extension to your browser’s toolbar for easy access. Now that you have the extension up and running, it’s time to access your LastPass web vault. To access it, go to the official LastPass website and click Log In. From there, enter your email and master password and you will be led to the LastPass web vault application. LastPass web vault. Image: Luis Millares Your LastPass vault is where all your passwords, logins and other credentials will be stored. 3. Use LastPass The main reason to use a password manager is to organize and secure your passwords. To illustrate how to do this, I’ve made a new account on LinkedIn. On LinkedIn’s signup page, you will see that there’s a LastPass icon on the username and password fields. Creating a LinkedIn account. Image: Luis Millares Clicking on the LastPass icon in the password field will bring up LastPass’ password generator. Here, LastPass automatically generates a random password for your new login. LastPass password generator. Image: Luis Millares At default, LastPass will generate a 16-character password — which you can configure to have uppercase and lowercase letters, numbers, and symbols. LastPass allows you to create a password with as many as 99 characters. In this case, I selected the first suggestion from LastPass’ password generator. Clicking on the generated password will prompt LastPass to ask you if you want to save the new login. Click “Add password,” and there you go — you’ve just created and saved your very first LastPass login. LastPass saving new login. Image: Luis Millares This login can now be found within your vault and LastPass will automatically be filled in the next time you’re on LinkedIn’s login page. With that, we’ve just saved our very first login using LastPass! SEE: IT Leader’s Guide to Cybersecurity Awareness Training (TechRepublic Premium) LastPass frequently asked questions (FAQs) Is LastPass no longer safe? Due to its recent data breaches, I don’t consider LastPass a safe password manager to use in 2025. For context, LastPass was involved in two data breaches in 2022, where the threat actor gained unauthorized access to encrypted customer data. While the password manager service has since implemented changes to bolster its security posture, there’s simply no reason to put your data at risk. This is especially true given there are other top-tier password managers that have not been hacked or involved in similar breaches. Does LastPass have a free version? Yes, LastPass has a free version but has limited features compared to its paid counterpart. On LastPass Free, you can only access your password vault while using one device type at a time. This means you can only view your passwords on a computer or a mobile device, not both at the same time. You also don’t get LastPass’ One to Many Sharing feature (which allows login detail sharing with multiple people), encrypted file storage, advanced multi-factor authentication, and emergency access. All these features are available on any of LastPass’ paid plans. Check out LastPass Free vs Premium, for a more in-depth comparison between the two. How to add LastPass to Chrome? To add LastPass to Chrome, visit the Chrome Web Store, search for LastPass, and click on “Add to Chrome.” From there, LastPass will provide you steps on how to

If you’ve been scrolling social media over the past week, you may have noticed miniature action figure versions of friends, family, or colleagues neatly wrapped in a blister pack. These plastic-fantastic portraits are the latest AI-powered photo trend to sweep the internet — especially LinkedIn. After digital avatars and Studio Ghibli-inspired selfies, we now have the action figure, produced using ChatGPT’s free image generator.  It’s all fun and games, right? But look closer, and behind the gloss and giggles lies some pretty crucial fine print. With this action figure, sustainability is not included.  Mel Morris, the founder of research engine Corpora.ai and former chairman of Candy Crush creator King, has slammed the environmental costs. “The LinkedIn action figure trend and its demand on GPUs is obviously unsustainable,” he said.  The 💜 of EU tech The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now! Some researchers estimate that using AI to generate text consumes 20-30 times more energy than a traditional search, depending on the model. Creating AI-generated images from written prompts uses at least double that amount, according to data from HuggingFace.  “GPUs aren’t infinite,” said Morris. “They come at a carbon cost, and this kind of casual overuse shows how disconnected we’ve become from the true costs associated with these tools.” Energy use from data centres, including for AI applications, is predicted to double over the next five years to 3% of global energy use, according to the International Energy Agency’s latest figures. Globally, almost half of that power is predicted to come from burning fossil fuels like coal and natural gas.   “As a technologist, I believe in the power of AI, but not at any cost,” said Morris. “If we want sustainable innovation, then we need to learn what genuine progress looks and feels like — I’m willing to bet that it’s not the dopamine hit after posting your AI-made figurine.” AI and sustainability feature heavily on the agenda of TNW Conference, which takes place on June 19-20 in Amsterdam. Tickets for the event are now on sale. Use the code TNWXMEDIA2025 at the check-out to get 30% off the price tag. source

By Christopher Cole ( April 11, 2025, 5:41 PM EDT) — An Ohio Republican has introduced House legislation to use some of the funds from the $42.5 billion Congress set aside for broadband expansion in 2021 to help defray the costs of obtaining satellite or fixed wireless broadband equipment and service…. Law360 is on it, so you are, too. A Law360 subscription puts you at the center of fast-moving legal issues, trends and developments so you can act with speed and confidence. Over 200 articles are published daily across more than 60 topics, industries, practice areas and jurisdictions. A Law360 subscription includes features such as Daily newsletters Expert analysis Mobile app Advanced search Judge information Real-time alerts 450K+ searchable archived articles And more! Experience Law360 today with a free 7-day trial. source

Amazon Nova Canvas is a foundation model for developers to create high-quality images. Image: Amazon Amazon is the latest tech giant to unveil a voice AI model. According to Amazon, its Nova Sonic is “a new foundation model that unifies speech understanding and speech generation into a single model, to enable more human-like voice conversations in AI applications.” Nova Sonic will compete with similar AI models by OpenAI, Google, and other tech companies. Nova Sonic understands more than words The Nova Sonic doesn’t just understand the speaker’s words, but it can also process the tone, style, and pace. The AI voice generator adapts to the conversation context, so dialogue flows more naturally, compared to the more stilted models from the first generations of Alexa. The Nova Sonic can do this because it combines multiple speech processing and generating functions into a single AI model instead of using multiple different models. Traditionally, AI voice tools involved running multiple models in sequence: a speech recognition model would convert speech to text, then a large language model (LLM) would process the input text and generate responses, and finally a text-to-speech model would convert text back to audio. This complex pipeline often stripped away the tone, style, and pacing of the speaker’s original dialogue. Since the Nova Sonic combines all of this in one model, it can adapt to the acoustic context of the input speech. It also responds more naturally to the cadences of human speech; for instance, it won’t interrupt when the speaker hesitates or pauses to take a breath. How to get Nova Sonic Nova Sonic is currently available via a new API in Amazon Bedrock, the company’s enterprise application building platform, and will simplify the development of voice applications. What developers need to know about Amazon Nova The tech giant recently introduced Amazon Nova Act, a new AI model trained to perform actions within a web browser. In addition, there is an Amazon Nova SDK for developers to explore. One of the foundation models is Nova Canvas for generating high-quality images; there are also models for generating text from different modalities as well as videos from text and image input. source

By Craig Clough ( April 14, 2025, 11:47 PM EDT) — A divided U.S. Securities and Exchange Commission refused Monday to redo settlements it inked with 16 financial firms over their failure to keep records of so-called off-channel communications, finding the “settlor’s remorse” the firms are suffering because others received better terms is not reason enough to modify their deals…. Law360 is on it, so you are, too. A Law360 subscription puts you at the center of fast-moving legal issues, trends and developments so you can act with speed and confidence. Over 200 articles are published daily across more than 60 topics, industries, practice areas and jurisdictions. A Law360 subscription includes features such as Daily newsletters Expert analysis Mobile app Advanced search Judge information Real-time alerts 450K+ searchable archived articles And more! Experience Law360 today with a free 7-day trial. source

U.S. President Donald Trump. Image: Gage Skidmore/Creative Commons U.S. President Donald Trump and his team have stated that the reciprocal tariff exemptions on electronic products are temporary. The administration excluded smartphones, computers, semiconductors, and other tech items from the 125% tariff on Chinese imports on April 11. Trump said his team is taking a look at the “whole electronics supply chain” as part of upcoming tariff investigations in a post shared on Truth Social. “NOBODY is getting “off the hook” for the unfair trade balances, and non monetary tariff barriers, that other countries have used against us,” he wrote. “Especially not China which, by far, treats us the worst!” This sentiment was echoed by Commerce Secretary Howard Lutnick who, on April 12, said that specific semiconductor tariffs would be coming in one or two months’ time. “These are things that are national security that need to be made in America,” he said in an interview with ABC. SEE: Were the White House’s Tariffs Calculations Done By AI? White House signals flexibility for key tech firms On April 13, Trump confirmed to reporters that he would be announcing the bespoke semiconductor tariff in the “very near future,” according to CNN, similar to other product-specific tariffs on aluminium, steel, and auto components. But he also said that some “flexibility” would be granted to specific companies in the sector, such as Apple. Exclusions to Trump’s tariffs for certain tech products were revealed in updated guidance from the US Customs and Border Control. It confirmed that the likes of memory cards, hard drives, motherboards, and servers would be excluded from the 10% global baseline tariff nor to the larger reciprocal tariffs imposed on countries with which the US has a trade deficit. China’s Ministry of Commerce described the exemptions as “a small step toward correcting its erroneous unilateral practice of ‘reciprocal tariffs’,” in a statement to state news outlet China Daily. However, tech products from China are still subject to the 20% tariff imposed to penalise Beijing for its failure to curb the export of fentanyl, as confirmed by White House Deputy Chief of Staff on Policy Stephen Miller on X. Karoline Leavitt, the White House press secretary, has indicated that the exemptions are to provide temporary relief to tech companies as they transition manufacturing operations to the States. “These companies are hustling to onshore their manufacturing in the United States as soon as possible,” she said in a statement to the BBC. Experts skeptical over short-term exemption impact Nevertheless, experts say a few weeks of exemptions are unlikely to make a difference when it comes to onshoring semiconductor production. “While chip manufacturing in the US is announced, those plants will not start delivering before 2027,” Mark Moccia, VP and Research Director at Forester, told TechRepublic in an email. He added: “US CIOs should plan to boost AI project budgets, as previously allocated budgets will likely not suffice to continue these projects at their current pace.” Trump’s tariffs have damaged tech giants’ value, and are likely to result in price hikes for American consumers Reciprocal tariffs for a number of countries were set to take effect on April 9, but Trump granted a 90-day reprieve for most of them. This was expected to have a positive impact on the stock value of the likes of Apple and NVIDIA, which took heavy blows when the tariffs were first announced. However, he imposed a 104% tariff on Chinese goods, later raising it to 145%. Beijing has retaliated with a 125% tariff on imports from the US, but will “ignore” any further hikes as they have “no practical economic significance,” according to its Commerce Ministry. These tariffs are expected to significantly impact Apple, which manufactures most of its devices in China, as do most other tech and AI companies. Tech analyst Dan Ives estimated a Chinese-made device that sold for about $1,000 would cost $3,500 if made in the US, a price point that many consumers may be unable to afford. source

Motorheads have long theorised that the extreme downforce generated by high-performance cars could one day let them drive upside down. Now, British carmaker McMurtry has turned that wild idea into a reality for the first time. In a stunt that would make Batman jealous, McMurtry drove its insanely fast electric Spéirling flipped over. More impressively, it did that for over a minute while keeping the vehicle completely stationary — save for a quick acceleration to prove the car wasn’t tied down.  While it might look like the car has defied gravity, in reality, it’s basic physics coupled with some clever engineering.  To drive upside down using conventional aerodynamics, an F1-style car would need to hit at least 100–150 mph to generate enough downforce to exceed its own weight and “stick” to the ceiling.   From Shark Tank to Tinder Swindler TNW Conference 2025 combines the latest breakthroughs in tech, the startup ecosystem & enterprise innovation But the Spéirling is a fan car. Twin electric turbines positioned behind the cockpit of the vehicle pull air from under the chassis and expel it through a rear-mounted exhaust system, creating a low-pressure zone underneath that presses the car onto the road. This system means the Spéirling can generate 2,000kg of downforce on demand. That’s great for grip in high-speed turns, but it also means the 1,000kg car can hang upside down, even when standing still.  That’s how McMurtry’s co-founder Thomas Yates was able to drive the Spéirling onto a rotating rig, flip it 180 degrees upside down, and have it stay there. No wires. No magnets. Just an insane amount of downforce.  The route to public roads for McMurtry McMurtry steered the Spéirling up a ramp and drove it upside down. Credit: McMurtry While this was just a stunt — for now — one can’t help but wonder what the future might hold. Stuck in tunnel traffic? Just drive up the wall, flip upside down, and cruise on the ceiling. Goodbye, gridlock.  The Spéirling Pure, the company’s first commercially available vehicle, is set to go on sale next year for £895,000. So, technically, you could try upside-down driving yourself.  “That said, customers are strongly advised not to try [driving upside down] at home,” a McMurtry spokesperson told TNW. If you’re mad enough to give it a go, McMurtry recommends contacting the company directly to discuss “pre-flight checks, preparation and safety equipment, and controlled demonstration environments.” For now, though, McMurtry is preoccupied with smashing records the right side up. On the same day as the Batmobile-like stunt, the Spéirling beat the all-time Top Gear track record by 3.1 seconds, dethroning a Renault F1 car at the top of the leaderboard. The McMurtry Spéirling’s fan-based tech gives it incredible grip in corners. It’s no slouch, either. Dual electric motors — one for each rear wheel — deliver over 1,000 horsepower combined. Paired with a featherweight chassis, this allows the car to clock 0–60mph in 1.5 seconds. source

After the stock market and hopes of consumer electronics companies such as Apple were buoyed by news of a tariff reprieve on consumer electronics on Friday, they were rapidly dashed by reports that no such exception was in the offing. Instead, these categories were moved to a different “bucket.” Businesses thrive on stability because they plan around rules of engagement — plans that entail commitments of significant time, resources, and capital expenditure. Ergo, markets perform better when businesses are confident that the rules are really the rules. When Friday’s policies are thrown out with Sunday’s brunch leftovers, companies will resort to one primary strategy: Do as little as possible and thereby do no harm. This is exactly what we recommend in our report, Consumer Marketing, CX, And Digital Leaders: How To Thrive Through Volatility (US): Times of extreme volatility often spawn organizationwide crises of “Everything must change, all at once” — a narrative that is both false and dangerous. Instead, keep a cool head, resist knee-jerk reactions, and fine-tune strategies precisely and creatively to adapt to only the meaningful and substantial changes in the market and business environment. What does Apple do on Monday now that Friday’s rules don’t apply? Here’s what I predict: Lobby the US administration hard and keep hopes alive for an exemption, for itself or the category. It’s been a successful route in the past and is really the best option in this tariff regime. Stay put and reduce risk, avoiding significant changes and accumulating inventory, as the company has been doing, to serve as an insurance policy (and if Apple doesn’t need to use it, that’s even better). Continue business-as-usual supply chain diversification, mostly independent of the current tariff volatility. The company has been moving production to countries such as India and Vietnam well before this administration; it can incorporate options to expedite this if the tariffs remain. Have Plan B pricing that looks at different price elasticities by product line. Come up with the right price points and offers (such as trade-in) that do not jeopardize upgrade cycles into more expensive phones and the adoption of the lucrative Apple ecosystem. Manage the China relationship, because the country matters not just as a supply source but as a market that is already under threat and can be a source of significant revenue pain. The trade war has added strain to the relationship, but Apple has to work hard to avoid blowback. To better manage your brand and business through this period of uncertainty and shifting consumer behaviors, please read our report, Consumer Marketing, CX, And Digital Leaders: How To Thrive Through Volatility (US). If you are a Forrester client, stay tuned for additional research on how CMOs can better manage uncertainty and volatility. Go to my Forrester bio and click “Follow” to be notified. Also, as a client, you can schedule time with me for an inquiry or guidance session or talk to your account team about workshops and strategy days on planning through uncertainty. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More It started with the announcement of OpenAI’s o1 model in Sept. 2024, but really took off with the DeepSeek R1 release in Jan. 2025. Now, it seems that most major AI model providers and trainers are in a new race to deliver better, faster, and cheaper “reasoning” AI language models — that is, ones that maybe take a little longer to respond to a human user, but ideally do so with better, more comprehensive, more well “reasoned” answers, which these class of models get by performing “chain-of-thought,” reflecting on their own conclusions and interrogating them for veracity before responding. ByteDance, the Chinese web media giant parent of TikTok, is the latest to join the party with the announcement and publication of the technical paper behind Seed-Thinking-v1.5, an upcoming large language model (LLM) designed to advance reasoning performance across both science, tech, math, and engineering (STEM) fields and general-purpose domains. The model is not yet available for download or use, and it’s unclear what the licensing terms will be—whether it will be proprietary/closed source, open source/free for all to use and modify at will, or somewhere in between. However, the technical paper provides some noteworthy details that are worth going over now and in advance of whenever they are made available. Built atop the increasingly popular Mixture-of-Experts (MoE) architecture Like Meta’s new Llama 4 and Mistral’s Mixtral before it, Seed-Thinking-v1.5 is built using a Mixture-of-Experts (MoE) architecture. This architecture is designed to make models more efficient. It essentially combines the capabilities of multiple models into one, each specializing in a different domain. In this case, the MoE architecture means that Seed-Thinking-v1.5 uses only 20 billion of the 200 billion parameters at a time. ByteDance says in its technical paper published to GitHub that Seed-Thinking-v1.5 prioritizes structured reasoning and thoughtful response generation. The results nearly speak for themselves, with Seed-Thinking-v1.5 outperforming DeepSeek R1 and approaching Google’s newly released Gemini 2.5 Pro and OpenAI’s o3-mini-high reasoner on many third-party benchmark evaluations. It even exceeds those two in the case of the ARC-AGI benchmark, which measures progress towards artificial general intelligence, seen as the goal or “Holy Grail” of AI. This model outperforms humans on most economically valuable tasks, according to OpenAI’s definition. Positioned as a compact yet capable alternative to larger state-of-the-art models, Seed-Thinking-v1.5 achieves competitive benchmark results. It introduces reinforcement learning (RL) innovations, training data curation and AI infrastructure. Performance benchmarks and model focus Seed-Thinking-v1.5 shows strong performance on a suite of challenging tasks, scoring 86.7% on AIME 2024, 55.0% pass@8 on Codeforces and 77.3% on the GPQA science benchmark. These results place it close to or matching models like OpenAI’s o3-mini-high and Google’s Gemini 2.5 Pro on specific reasoning metrics. On non-reasoning tasks, the model was evaluated through human preference comparisons and achieved an 8.0% higher win rate over DeepSeek R1, suggesting that its strengths generalize beyond logic or math-heavy challenges. To address saturation in standard benchmarks like AIME, ByteDance introduced BeyondAIME, a new, harder math benchmark with curated problems designed to resist memorization and better discriminate model performance. This and the Codeforces evaluation set are expected to be publicly released to support future research. Data strategy Training data played a central role in the model’s development. For supervised fine-tuning (SFT), the team curated 400,000 samples, including 300,000 verifiable (STEM, logic and coding tasks) and 100,000 non-verifiable problems like creative writing and role-playing. For RL training, data was segmented into: Verifiable problems: 100,000 rigorously filtered STEM questions and logic puzzles with known answers, sourced from elite competitions and expert review. Non-verifiable tasks: Human-preference datasets focused on open-ended prompts, evaluated using pairwise reward models. The STEM data leaned heavily on advanced mathematics, accounting for over 80% of the problem set. Additional logic data included tasks like Sudoku and 24-point puzzles, with adjustable difficulty to match model progress. Reinforcement learning approach Reinforcement learning in Seed-Thinking-v1.5 is powered by custom actor-critic (VAPO) and policy-gradient (DAPO) frameworks, developed to address known instabilities in RL training. These techniques reduce reward signal sparsity and enhance training stability, especially in long chain-of-thought (CoT) settings. Reward models play a critical role in supervising RL outputs. ByteDance introduced two key tools: Seed-Verifier: A rule-based LLM that checks if generated and reference answers are mathematically equivalent. Seed-Thinking-Verifier: A step-by-step reasoning-based judge that improves judgment consistency and resists reward hacking. This two-tiered reward system enables nuanced evaluation for both straightforward and complex tasks. Infrastructure and scaling To support efficient large-scale training, ByteDance built a system atop its HybridFlow framework. Execution is handled by Ray clusters, and training and inference processes are co-located to reduce GPU idle time. The Streaming Rollout System (SRS) is a notable innovation that separates model evolution from runtime execution. It accelerates iteration speed by asynchronously managing partially completed generations across model versions. This architecture reportedly delivers up to 3× faster RL cycles. Additional infrastructure techniques include: Mixed precision (FP8) for memory savings Expert parallelism and kernel auto-tuning for MoE efficiency ByteCheckpoint for resilient and flexible checkpointing AutoTuner for optimizing parallelism and memory configurations Human evaluation and real-world impact To evaluate alignment with human-centric preferences, ByteDance conducted human testing across a range of domains, including creative writing, humanities knowledge and general conversation. Seed-Thinking-v1.5 consistently outperformed DeepSeek R1 across sessions, reinforcing its applicability to real-world user needs. The development team notes that reasoning models trained primarily on verifiable tasks demonstrated strong generalization to creative domains—an outcome attributed to the structure and rigor embedded in mathematical training workflows. What it means for technical leaders, data engineers and enterprise decision-makers For technical leads managing the lifecycle of large language models—from data curation to deployment—Seed-Thinking-v1.5 presents an opportunity to rethink how reasoning capabilities are integrated into enterprise AI stacks. Its modular training process, which includes verifiable reasoning datasets and multi-phase reinforcement learning, particularly appeals to teams looking to scale LLM development while retaining fine-grained control. ByteDance’s moves to introduce Seed-Verifier and Seed-Thinking-Verifier offer