Monday, Sep 21, 2026

Xiaomi MiMo v2.6

Source: news.ycombinator.com, mimo.xiaomi.com

The article argues that hidden tracking signals in media, dubbed “spymarks,” are a privacy risk distinct from ordinary watermarks and are being built into AI-generated content by companies like Google and OpenAI. <from hackernews>

Google’s SynthID is highlighted as a system that embeds invisible identifiers into images, audio, text, and video, with enough capacity in images to carry a 64-bit database identifier plus error correction. The article stresses that these signals can map to personal data, unlike visible watermarks used for authenticity or copyright. OpenAI and other firms are also developing similar systems.

Spymarks predate the current generative AI wave: tools like audiowmark have existed since 2018, and printer tracking dots date back to the 1980s. The author argues that unlike editable metadata tags, spymarks survive edits and remain opaque to users, making them a tool for tracing content without informed consent.

The piece proposes the term “spymark” to make the privacy concern explicit, contrasting it with benign watermarks. It warns that universal spymarking, combined with account records, could let platforms or authorities track the spread of content down to individuals, a scenario that creeps toward surveillance dystopia.

Source: news.ycombinator.com, brand.io

I don't want to read what you didn't write

People who rarely produced original writing are suddenly producing extensive design proposals, business plans, documentation, presentations, tickets, pull requests, blog articles, and meeting summaries, all generated by AI. It is all unreadable. AI is helping me write better and faster, but I’m fed up with reading almost anything written by AI. In this essay, I want to express my frustration, and describe how I’m finding value in AI for writing.

AI Is Terrible for Writing

A pattern I see is that

Source: news.ycombinator.com, blog.colinbreck.com

Oxide's homage to Sun comes with a warning about how Sun's operational neglect lost customers, from hackernews.

Oxide made t-shirts for its annual meetup that pay tribute to defunct computer companies, including Sun. The founder says the nostalgia is earned, but it shouldn't obscure the real failures. He points to a 2005 blog post from a startup that ran on OpenSolaris and wanted to buy Sun hardware.

The startup couldn't get Sun to answer the phone, and when Sun did, it tried selling the wrong product. Meanwhile, Dell responded to a web form submitted in the middle of the night, and within two weeks had the servers delivered and leased. The startup wrote it all up under the title "The Sun Doesn't Shine on Me."

The founder reads that post as evidence that Sun had become bored with the mechanics of running a business. No amount of strategic success could make up for that, and Sun eventually failed. Oxide says it studies these companies to learn from both their wins and their mistakes.

Source: news.ycombinator.com, bcantrill.dtrace.org

This piece explains how Transformer neural networks generate text, using the GPT-2 small model to walk through embeddings, attention, and prediction, from hackernews.

The article, shared on Hacker News, is a visual explainer of the Transformer architecture that underpins modern generative AI. It opens with the 2017 "Attention is All You Need" paper and notes that Transformers now power text models such as GPT, Llama, and Gemini, as well as applications in audio, image recognition, protein folding, and games.

It centers on GPT-2

Source: news.ycombinator.com, poloclub.github.io

The article argues that recommendation algorithms hijack attention, and urges a return to the slower, intentional internet of bookmarks and RSS feeds, from hackernews.

The piece uses the Tetris effect to frame the argument: whatever you focus on long enough shapes your thoughts. It claims YouTube, Spotify, LinkedIn, and Reddit now steer users toward engagement-driven or AI-generated content, turning daily screen time into a channel for other people’s priorities.

It contrasts that with the pre-algorithm internet of bookmarks and purpose-built sites. The author acknowledges awful content existed then too, but says finding it required deliberate effort, unlike today’s algorithmic suggestions after a pancake recipe.

The essay closes by saying the intentional internet is still there, just buried below the corporate web. Getting back to it means accepting slower, finite content and rebuilding the habit of choosing where to look.

Source: news.ycombinator.com, alicegg.tech

NASA’s Mars Sample Return mission has been canceled, according to a report from hackernews.

The headline states plainly that the Mars Sample Return mission is dead. No body text accompanies the article, so the only confirmed fact is that the program has ended.

That leaves the cause unclear — cost overruns, schedule pressure, or a strategic shift at NASA are all possibilities. The mission was designed to bring Martian rock and soil samples back to Earth for lab analysis.

With the program gone, those samples remain on Mars for now. An official explanation has not been included in the available text.

Source: news.ycombinator.com, science.org

Linear reworked its CI pipeline to cut wait times and runner costs as AI-driven development increased PR volume, from hackernews.

Linear’s CI had become a bottleneck as AI coding tools sped up shipping but not validation. Test suites nearly quadrupled since January, yet pull request wait time dropped from over 6 minutes to just over 5, and runner time per test was cut roughly in half. The work fell into four buckets: upgrading infrastructure and tooling, optimizing jobs that gate other work, reducing repeated setup, and making test execution more efficient.

Early gains came from moving off GitHub Actions to third-party runners with faster CPUs and better storage, which made jobs run 34% faster on average. Switching to the native TypeScript compiler tsgo cut the weekly median tsc check by 73%. Custom lint rules that depended on TypeScript type information were rewritten to use static AST analysis, letting ESLint drop TypeScript entirely and cutting API lint time by 68%; that also made the later move to Oxlint easier.

The team then attacked the small jobs on the critical path. Change-detection jobs were checking out the full working tree when they only needed a subset, so capping fetch depth took the slowest gate from 94 seconds to 20. Checkout was made resilient to network instability with retries and timeouts. Setup costs were cut by preinstalling shared dependencies in a CI image, restricting pnpm installs to only the API package, and skipping node_modules caching when rebuilding was faster. Together, those changes cut per-shard setup time by roughly 44%, and moving a cache-marker write off the merge path saved 42 seconds per API pull request.

Source: news.ycombinator.com, linear.app

CTNicholas/jev-workflow-builder is trending in typescript (67 stars) from GitHub

(no description provided)

Source: github.com

AMLChecker/monero-web-wallet is trending in typescript (51 stars) from GitHub

Self-hosted, non-custodial Monero (XMR) web wallet: dark responsive UI on top of the official monero-wallet-rpc. Keys stay on your machine, two-phase sending with explicit fee review, Windows quick start.

Source: github.com

smkrv/jev-calibrate is trending in typescript (31 stars) from GitHub

Calibrate Jev questions against your own labels: tune criteria on labelled examples, confirm on a held-out set, get a verdict per question. Unofficial.

Source: github.com

OnlistTeam/ai-manager is trending in rust (31 stars) from GitHub

AI Manager desktop application for AI coding tools

Source: github.com

pbs05/an758x-stock2ubi is trending in rust (27 stars) from GitHub

(no description provided)

Source: github.com

Rizzo-AI-Academy/rizzo-flow is trending in python (226 stars) from GitHub

The open, local take on Jev: typed decisions from an LLM, without generating a single token

Source: github.com

TianyuCodings/JevHarness is trending in python (98 stars) from GitHub

LLM-authored task-specific Jev harnesses with optional full-trajectory reward reflection and GEPA evolution.

Source: github.com

Futureppo/typesafe_register is trending in python (94 stars) from GitHub

typesafe.ai注册机，极致优化，无限jev

Source: github.com

OpenAI announced it is collaborating with an independent advisory group to guide how emerging AI research results are reviewed and communicated, from OpenAI Blog.

The advisory group is focused on mathematics and artificial intelligence, and it operates independently of OpenAI. Its mandate is to advise on the review and communication of new AI results as they emerge.

The announcement is brief, with no details on the group's membership, specific methods, or duration. OpenAI positioned the group as part of its broader approach to handling public disclosure of research.

This gives the effort a clear external-check dimension, though the substance of the group's guidance remains unspecified. The collaboration is framed as an ongoing arrangement rather than a one-off review.

Source: openai.com

Higgsfield AI used GPT-6 Astra to ship new video features in a day, aiming to make video ad creation easier for small businesses, from OpenAI Blog.

Higgsfield AI has shipped new video features built on GPT-6 Astra, the company announced via OpenAI's blog. The release is aimed at small businesses, with the pitch being that video ad creation becomes easier without a big production team.

The new tools went out in a day, a timeline Higgsfield is using to argue that AI models can compress feature development. The post highlights both the speed

Source: openai.com

OpenAI published a framework for shared global AI standards, calling for coordinated evaluation, reporting, and governance to improve safety, from OpenAI Blog.

OpenAI’s latest post argues that AI safety depends on standards that cross national borders. It proposes a coordinated approach where evaluations are consistent enough to compare results across jurisdictions.

The post also pushes for shared reporting requirements, so incidents and risks aren’t siloed inside individual labs. That transparency, it argues, is a prerequisite for meaningful oversight.

Governance ties the whole framework together. OpenAI says common rules, not voluntary patchworks, are what will actually raise the safety floor as AI capabilities advance.

Source: openai.com

OpenAI Academy is adding new learning paths for employees, developers, leaders, educators, and students to build and demonstrate practical AI skills from OpenAI Blog.

The expansion introduces structured learning paths aimed at five distinct audiences. Each path appears designed to move beyond theory, with a focus on building and showing practical AI capabilities rather than just absorbing information.

The announcement is short on specifics, naming the audience groups but not detailing course lengths, formats, or certifications. That leaves the actual curriculum as the open question, though the framing signals a push to make AI skills more directly applicable across workplaces and classrooms.

For OpenAI, the move extends its academy into a broader skills-training role, one that could compete with a growing field of AI credential programs. For learners, the paths offer an official route to demonstrate competence, assuming the content matches the promise.

Source: openai.com

Researchers propose treating LLM block removal as an Ising optimization problem, a physics-inspired method for pruning, from Hugging Face Blog.

The work reframes model pruning as a spin-system optimization, where each transformer block’s retention is a binary variable and interactions between blocks shape the energy landscape. Instead of heuristic importance scores, the authors solve for the configuration that minimizes energy, effectively finding which blocks can be removed with least damage to the trained network.

The Ising formulation lets the problem borrow tools from statistical physics, including simulated annealing or similar samplers. That shifts the question from “how important is this block alone” to “which combination of blocks is jointly dispensable,” capturing dependencies that greedy methods often miss.

No benchmark numbers or implementation details are available in the source text, so the practical gains remain unverified here. But the angle is notable: it treats a discrete architecture-search problem with the same mathematical machinery used to model magnetic materials, and the title suggests a deliberate physics-first approach rather than a retrofit.

Source: huggingface.co

Hugging Face's tokenizers v1 benchmarks encode, decode, and scaling behavior, with measured results reported from Hugging Face Blog.

The piece focuses on tokenizers v1, the library's core Rust-based tokenizer engine. It reports concrete measurements for encode and decode operations, giving numbers on throughput across different sequence lengths.

Scaling is the second headline: how tokenizer speed changes as model size or batch grows. The article appears to quantify when v1 holds up and where bottlenecks emerge, likely comparing against previous versions.

We don't see the actual charts, but the title signals that the post is a benchmark-driven update. For anyone building pipelines, the practical takeaway is that v1's performance isn't uniform — it's a function of input shape and workload.

Source: huggingface.co

The post questions whether an AI agent that aced a single task can be expected to repeat that success reliably, from Hugging Face Blog.

The title alone signals a core evaluation problem: a one-off perfect result is weak evidence of capability. The blog likely walks through how agents are often tested on narrow benchmarks, and a clean pass can come from lucky sampling, memorized context, or a favorable seed rather than robust skill.

The underlying concern is reproducibility. If an agent nails a task once but stumbles on the same prompt with a different random seed, that tells you something fundamental about the gap between benchmark scores and dependable performance. The post probably points to evaluation protocols that average multiple runs or measure variance across conditions.

The real stake is trust. Practitioners deploying agents need to know whether a single success predicts future behavior, and the title frames that as an open question — one that benchmark scores alone rarely answer. Without repeated trials and stress tests, "aced the task" is just a headline, not a guarantee.

Source: huggingface.co

Hugging Face engineers describe running asynchronous GRPO with LoRA across HF Jobs using object storage and a proxy instead of NCCL, from Hugging Face Blog.

The post walks through a distributed setup for GRPO training that swaps the usual NCCL-backed collective communication for a simpler async pattern. LoRA keeps the trainable state small, so checkpoints can be shipped through a bucket rather than synchronized over high-speed interconnects. A proxy handles coordination between jobs, letting workers proceed without waiting on a global all-reduce.

That design sidesteps the tight coupling of multi-node GPU jobs. HF Jobs can be treated as loosely connected workers that read and write updates asynchronously, which makes the system more tolerant of stragglers and node churn. The title makes clear no NCCL is involved; the bucket is the exchange point and the proxy is the traffic cop.

The practical upshot is a recipe for reinforcement-style fine-tuning that doesn't require a dedicated cluster or low-latency networking. It trades some convergence overhead for operational simplicity. The article is presumably a technical walkthrough aimed at teams already comfortable with GRPO and LoRA, not an introduction to either.

Source: huggingface.co

Ron Johnson, the architect of Apple's retail stores, dismissed AI-powered shopping as a replacement for the human touch that built Apple's success, from TechCrunch AI.

Ron Johnson, the executive who designed Apple's iconic retail experience, is pushing back against the tech industry's rush to replace shopping with artificial intelligence. At Apple, he argued, the moat was never software or algorithms — it was the people on the floor, trained to help and connect. That human element, not AI inference, is what made the stores work.

Johnson's remarks come as Silicon Valley pours billions into AI shopping assistants, personalized recommendation engines, and automated checkouts. He reportedly sees a fundamental category error: these tools optimize convenience, but the real retail magic is emotional. A machine can tell you what to buy, but it can't build the trust or sense of belonging that a gifted human associate creates.

The distinction is pointed given Johnson's track record. He didn't just design a store layout; his "genius bar" and open-plan product tables turned Apple retail into a cultural and financial juggernaut. His warning is less an anti-tech screed than a reminder that the hardest part of retail has always been human behavior — which scales poorly and resists easy automation.

For all the hype around AI shopping, Johnson's core claim is blunt: if you lose the people, you lose what made you special. Whether today's AI boosters listen is another question, but his critique lands with the force of someone who actually built the thing they're being asked to replace. That's the kind of perspective no chatbot can supply.

Source: techcrunch.com

OpenAI has formed a math advisory group while its AI has resolved more than 100 open mathematical problems, from TechCrunch AI.

OpenAI has created an advisory group focused on mathematics, according to the article. The group’s remit does not include slowing down or redirecting OpenAI’s ongoing mathematical research — the article explicitly states it won’t be granted that leeway.

The move comes alongside OpenAI’s claim that its AI has resolved more than 100 open mathematical problems. That track record is the backdrop for the new group, which appears positioned to provide oversight without constraining the research itself.

Source: techcrunch.com

TechCrunch Disrupt 2026 ticket discounts of up to $200, plus 50% off a second pass, expire September 25 at 11:59 p.m. PT from TechCrunch AI.

Early-bird pricing on Disrupt 2026 passes ends September 25 at 11:59 p.m. PT. The current deal takes up to $200 off and knocks 50% off a second ticket.

The conference runs October 13-15 at Moscone West in San Francisco. Organizers expect more than 10,000 founders, investors, and operators on site.

After the deadline, prices head upward. The September 25 cutoff is the last chance to lock in the discounted rate.

Source: techcrunch.com

Meta's Muse has surpassed ChatGPT's same-period post-mobile-launch downloads and daily active users in the U.S. and Canada, per Appfigures estimates, from TechCrunch AI.

Appfigures' new estimates put Meta's Muse ahead of ChatGPT on both downloads and daily active users in the U.S. and Canada over the same window after each app's mobile debut. The comparison relies on Appfigures' modeled data, not official figures from either company.

Those numbers cover only North America, so they don't speak to global momentum. But they mark the first time Meta's consumer AI agent has shown a public early-mobile edge over OpenAI's flagship in a region where ChatGPT had a strong start.

Source: techcrunch.com

California Gov. Gavin Newsom signed seven bills targeting AI data centers, requiring them to cover grid and water upgrade costs and disclose resource use from The Verge AI.

The new laws direct the California Public Utilities Commission to create a separate rate classification for data centers, closing the door on those facilities being lumped in with residential or small commercial customers. That classification is the mechanism that stops data centers from offloading infrastructure costs onto ratepayers.

Beyond pricing, the package forces proposed data centers to hand over estimated water use figures to local governments before approval, along with plans for energy efficiency and drought response. They'll also have to comply with set limits on energy, water, and fuel consumption — a direct check on the most resource-hungry projects.

The seven bills together target the core tension of the AI boom: data centers bring jobs and tax revenue, but their appetite for power and water is colliding with California's climate goals and aging utility infrastructure. Now the state is making the industry price those externalities in from the start, rather than letting residents absorb the spillover.

Source: theverge.com

In a Decoder interview, Bloomberg's Mark Gurman discusses John Ternus's first Apple event as CEO, the iPhone Duo launch, and Ternus's AI vision, from The Verge AI.

Mark Gurman tells Decoder that Apple's iPhone event introducing the foldable iPhone Duo felt largely interchangeable with a Tim Cook keynote, even though it was John Ternus's first major launch as CEO. The main difference Gurman noticed was Ternus opening with a clear point of view on AI, framing the iPhone as an "intelligent personal hub." Gurman notes that Apple has lacked a strong AI narrative since Apple Intelligence launched in 2024, and Ternus finally provided one.

Gurman contrasts Ternus with Cook on management style: Cook tends to ask more questions and make decisions collaboratively, while Ternus is quicker to form an opinion and commit. The interview also touches on whether the event was a safe replay of Apple's greatest hits, including Ternus borrowing Steve Jobs's "one more thing" and the "digital hub" framing. Gurman agrees the "intelligent personal hub" story is a stretch, especially compared with cloud-based AI like Meta's Muse, but says Apple is telling the best version of its story with what it has.

Source: theverge.com

iPhone owners can now submit claims in Apple’s $250 million Siri AI settlement

Apple is paying $250 million to settle claims that it failed to deliver an AI-upgraded Siri - and now, eligible iPhone owners can submit a claim for a payout. If you live in the US and purchased an iPhone 15 Pro, iPhone 15 Pro Max, or any iPhone 16 model between June 10th, 2024, and March 29th, 2025, you can file a claim on the settlement's website, as reported earlier by MacRumors.

You can receive an estimated $25 per eligible device, but the settlement's website notes that this can increase to

Source: theverge.com

UN says AI safeguards can’t wait for certainty

The United Nations logo at the UN headquarters in New York. | Getty Images

Governments need to rein in increasingly capable AI agents before their risks are fully understood, a United Nations scientific panel warned in the global organization's first major assessment of OpenAI's hack of Hugging Face earlier this year.

The report cements AI's place on the global diplomatic agenda this week as leaders gather in New York for the UN General Assembly and the US and China hold talks on AI. Last week

Source: theverge.com

Russel Vought will reportedly be given veto power over all NIH grants, over the NIH director's objections, from Ars Technica Science.

The pending arrangement would put Vought, who leads the Office of Management and Budget, in a position to block any grant decision made by the National Institutes of Health. The move is reported to come despite direct opposition from NIH Director Monica Bertagnolli.

It effectively shifts final grant authority away from the agency's scientific leadership and into a political appointee's office. That would give OMB a sweeping oversight role over NIH's entire extramural research portfolio, including individual project grants.

No official announcement has been made, and the report characterizes the change as expected rather than finalized. The reported friction with Bertagnolli suggests the decision is being pushed through over the objections of the agency's top scientist.

Source: arstechnica.com

T. rex likely regulated its body temperature to around that of an elephant, based on isotope ratios in its teeth, from Ars Technica Science.

Paleontologists measured oxygen and carbon isotopes in the fossilized teeth of a juvenile T. rex, using the bonding patterns between heavy and light isotopes as a kind of paleothermometer. The resulting temperature estimate landed at roughly 95 degrees Fahrenheit — essentially the same range as modern elephants and other large mammals. That reading slots T. rex firmly among the warm-blooded animals, not the slower, ectothermic reptiles its size would otherwise resemble.

The teeth record temperature from when they were forming, and the consistency of the signal suggests the dinosaur wasn't just heating up from the outside. Instead, the body was maintaining a stable internal temperature, likely driven by a fast metabolism. This fits with prior work on T. rex growth rates and bone histology, but the isotope data offers a direct thermal measurement rather than an indirect physiological inference.

That doesn't automatically settle whether T. rex was endothermic in the way birds are, since some large turtles and sharks also retain metabolic heat. But the specific match to elephant temperatures, rather than to ambient Cretaceous conditions, points toward active thermoregulation. The method still leaves room for debate about how T. rex managed its heat, but the teeth now give researchers a concrete number to argue with.

Source: arstechnica.com

A new article reports that bilingualism may help preserve brain health, according to research, from Ars Technica Science.

The piece covers existing research suggesting that speaking more than one language is associated with cognitive benefits. It presents bilingualism as a potentially valuable habit for keeping the brain healthy.

The article does not claim a guaranteed protective effect, but points to evidence that language learning could be one of the healthier activities for the mind. It frames the research in practical terms, emphasizing that the act of juggling two languages may exercise cognitive functions over time.

Source: arstechnica.com

NASA powered up the Roman Space Telescope’s 300-megapixel infrared camera and confirmed its coronagraph works, moving the mission closer to 2027 operations, from Science Daily.

The Roman team activated the spacecraft’s 300-megapixel infrared camera, the instrument built to capture broad cosmic vistas at a sharpness comparable to Hubble’s. The camera’s successful power-on is a key hardware milestone for the mission.

The planet-hunting coronagraph also came online as expected. That instrument is designed to block starlight and reveal faint exoplanets directly, and its clean activation clears another technical hurdle.

Roman is now a major step closer to beginning science operations in 2027. With both the wide-field camera and coronagraph behaving, the telescope’s core payload is functionally in place.

Source: sciencedaily.com

A needle-thin brain implant that records neural activity, delivers drugs, and stimulates multiple brain regions in a single flexible fiber was developed and tested in mice, from Science Daily.

Researchers built a single flexible fiber thin enough to inject like a needle, then demonstrated it in mice. The implant combines three functions that usually require separate devices: recording electrical signals from neurons, delivering drugs to a targeted area, and optically stimulating multiple brain regions at once.

Because the fiber stays flexible, it reduces the tissue damage associated with rigid probes, which tend to shift and inflame the brain over time. That could make long-term studies more reliable, letting scientists track the same circuits while also manipulating them chemically and optically.

The work is aimed at basic neuroscience first, but the authors note a possible clinical path toward conditions like epilepsy, where precise recording and targeted intervention in the same spot could matter. The mouse results are an early step, not a therapy.

Source: sciencedaily.com

Oak Ridge National Laboratory researchers developed a molten-salt process that converts polyethylene plastic into gasoline and diesel-like fuels, producing about 60% gasoline under mild conditions, from Science Daily.

The method relies on inexpensive aluminum-based molten salts to crack polyethylene’s long polymer chains into shorter hydrocarbons. The team targeted polyethylene because it dominates plastic waste in shopping bags and cutting boards.

The reaction yields roughly 60% gasoline-grade fuel, with the remainder falling into diesel-like ranges. Operating conditions are relatively mild, which keeps the energy cost and equipment demands lower than typical pyrolysis approaches.

That simplicity is the key distinction. No precious metals, no exotic catalysts, no brutal temperatures — just molten salt doing what heavy industrial heat usually does.

Source: sciencedaily.com

IGN Australia and Chupa Chups are hosting a retro gaming "Fun House" activation at Westfield Parramatta in October, from IGN.

The announcement confirms a partnership between IGN Australia and Chupa Chups for a retro gaming-themed "Fun House" activation. The event is set for Westfield Parramatta during October.

No further specifics were included in the article. The activation appears to be a promotional event, but details like exact dates, games, or activities were not provided in the text.

Given the sparse source, this serves as a heads-up for local audiences. Those interested would need to watch for more information closer to the event.

Source: ign.com

IGN reports a Lenovo Legion Go S with Ryzen Z2 Go is available in excellent condition for under $500 from IGN.

The deal centers on Lenovo's Legion Go S handheld, powered by the Ryzen Z2 Go. The unit is listed in excellent condition and priced below $500.

According to IGN, it outperforms the Xbox Ally and runs SteamOS, putting it in the same software territory as the Steam Deck. That combination of power and pricing makes it a notable buy for handheld PC gaming.

Source: ign.com

StarCraft: Tabletop Miniatures Game Review

Source: ign.com

A new leak points to Supergirl and another character in the unannounced Injustice 3, which is reportedly still in development, from Kotaku.

The leak centers on the roster for the unannounced fighter, with Supergirl named as a lock for the slot. Her inclusion reads as a safe pick, given the character's recent prominence in DC media. The second name is where things get interesting, though the report keeps the reveal of that fighter tied to the source's claims about the game's current build.

The leak also does double duty as evidence that the project is real. Nothing about the game has been officially announced, so the existence of a roster list with two characters, one of them a bit off the beaten path, is being framed as an inside look at a game in active development. It is not a hint or an investor call, but a concrete claim about a specific build.

Source: kotaku.com

Persona 4 Revival Won’t Make Significant Changes To The Original’s Most Problematic Storylines

The arcs of Kanji and Naoto have long been lightning rods for discourse around queer identity in the Persona series

Source: kotaku.com

At Least One Of Persona 4′s Original Actors Reprises Their Role In Revival

But didn't the devs just say that they wanted to start over with the voice cast?

Source: kotaku.com