ContentBuffer
Nineteen jurors on the 50-plus-member Independent Games Festival panel just said they won't vote for any game using generative AI. Then House House, the studio behind Big Walk, said its game used LLMs for a bug fix. The fallout is messy, and it's a preview of how every studio will get judged. If you ship games or any software with AI-assisted tooling, this is your problem too. Valve doesn't ask for code or chatbot usage disclosure, and its form targets AI-generated content, not efficiency tools built into modern engines. The IGF has no AI distinction at all, which is why House House disclosed the fix when entering. The catch? That honesty is now the story. The numbers: AI touched well under 0.1% of Big Walk's code. The game holds a Very Positive Steam rating despite 400+ review bombs, since Steam's anti-bomb tools are dampening the damage. The IGF previously rescinded awards for Clair Obscur: Expedition 33 over AI-made textures, so precedent exists. Judge Liz Ryerson said the AI use wasn't disqualifying for her, and Dominic Tarason called punishing a bug fix absurd. Hot take: a line has to be drawn somewhere, and a bug fix is not where it belongs.
A new open-source app called klad lets you paint and animate on a canvas shared with a coding agent. You decide what to draw, and the agent works alongside you on the same canvas. No image model is involved, which is the whole point. If you're building games, this is for you. klad targets props, characters, backgrounds, and frame-by-frame animation, and it can generate in-betweens and check arcs for realism. Exports go out as PNGs, sprite sheets, or a Godot SpriteFrames scene that keeps each clip's timing. The app also ships a command-line tool, so you can script drawings into your build. Every stroke and fill is stored as data, so you can recolour, bend, move, or delete them later. It runs on macOS, Linux, and Windows, needs no account, and checks for updates once a day, which you can turn off. The license is free for personal and commercial use as long as your company has under 500,000 euros in revenue. This is the first public version, so expect rough edges. Let's dive deeper…
In today's ContentBuffer update:
Nvidia's CEO Caught in Alleged Chinese Honeytrap and Chip Diversion Probe
AlphaGenome 2 Maps Genetic Variants Cell by Cell, With Nearly 2x the Hits
JioHotstar's six-way test: retrieval helps agents sometimes, hurts them other times
DeepMind's AI agents formally verify Signal's crypto in four weeks, no new bugs found
5 new AI tools & 5 new AI jobs
More tech news
Latest Development
Image source: cdn.arstechnica.net
Summary: Nineteen IGF jurors refused to vote for any game using generative AI, and they want a full ban next year. Big Walk's developer House House disclosed an AI-assisted bug fix touching under 0.1% of its code, and Steam reviewers responded with 400+ negative reviews. Expect more studios to audit their toolchains before submitting.
Key Points:
The 19 jurors sit on a panel of more than 50 members judging the Independent Games Festival Awards, and they want future rules to disallow any AI-generated work in entries.
House House says LLMs were used sparingly to help work through difficult technical challenges and bugs, representing well under 0.1% of Big Walk's total code.
Big Walk released on Steam earlier this year, and the AI-assisted fix went undisclosed because Valve does not request disclosure of code or chatbot usage from developers.
The IGF rescinded awards for Clair Obscur: Expedition 33 after its developers admitted some in-game textures were made with AI tools, setting a clear precedent for enforcement.
Big Walk's Steam page has taken more than 400 negative reviews, yet it still holds a Very Positive overall rating because Steam's anti-review-bomb tools are limiting the impact.
Why it matters: If your team ships on Steam, Valve's disclosure form only targets AI-generated content, so a Copilot-assisted fix in your engine code doesn't need to be declared. But festival submissions are a different game, and the IGF makes no distinction between AI-assisted fixes and AI-generated assets. Audit your build pipeline and tooling logs before entering any awards, because a single undisclosed LLM-touched commit can become a public controversy.
Image source: repository-images.githubusercontent.com
Summary: klad is a new open-source 2D painting and animation app built to share a canvas with coding agents. It exports sprite sheets and Godot SpriteFrames scenes, and it's free for companies under 500,000 euros in revenue. Early days, but game devs should clone the repo and try it.
Key Points:
Every brush stroke and fill is stored as editable data, so you can recolour, reshape, move, or delete them long after you drew them.
The app includes a brush editor, layers, masks, clipping, effects, and a timeline with onion skins and clips for frame-by-frame work.
Exports cover PNG files, sprite sheets, and a Godot SpriteFrames scene that preserves each clip's timing down to the frame.
Commercial use is free for companies with under 500,000 euros in revenue, and the source code is open for anyone to compile and modify.
The JSON-based file format lets other apps and engines read and write klad drawings, and the command-line tool scripts batch work.
Why it matters: If you're a solo indie or a small studio shipping 2D Godot games, klad turns sprite production into something your build pipeline can touch. The JSON file format and CLI mean a coding agent can read, edit, and export drawings without a human opening the editor. Teams under 500,000 euros in revenue get the whole thing free, which beats most paid sprite tools on price alone.
Image source: media.wired.com
Summary: Nvidia CEO Jensen Huang is tangled in a reported Chinese espionage probe involving an alleged intelligence asset and Megaspeed, a cloud firm that bought billions in Nvidia chips. The board says no secrets leaked, but Megaspeed's diversion risk is the part to watch.
Key Points:
US intelligence agencies warned top officials that Alice Huang was a Chinese intelligence asset, allegedly reporting to the Ministry of State Security intelligence agency.
Megaspeed's subsidiaries have purchased billions of dollars worth of Nvidia chips, making the cloud provider a central piece of the diversion concern.
A government audit requested last year found Megaspeed could divert chips to Chinese intelligence agencies, a finding that predates the current espionage reporting by months.
Nvidia's CEO has been publicly linked to a Megaspeed executive since at least 2024, according to reporting cited in the intelligence briefings and coverage.
Nvidia shares rose about 600 percent over five years but plummeted roughly 25 percent in the past month as the espionage story spread.
Why it matters: If you rent GPUs from neoclouds like Megaspeed, your training jobs run on hardware whose provenance is now under government scrutiny. Teams signing reserved capacity contracts should ask vendors for documented chip chain-of-custody and export compliance terms before renewing. Platform and procurement leads at AI startups should also track whether Washington tightens export rules again, because a sudden restriction could strand capacity you already paid for.
Image source: platform.theverge.com
Summary: Google DeepMind's AlphaGenome 2 predicts how genetic variants change gene expression in specific cells. It finds nearly twice as many expression-altering variants as its predecessor, trained on 100M+ cell profiles. If you do drug target work, the API is open for non-commercial research today.
Key Points:
Trained on over 100 million individual cell profiles spanning 68 tissues, a scale that only became feasible with newer single-cell capture technologies.
Identifies nearly twice as many variants that change gene expression in cells compared with AlphaGenome, its previous state-of-the-art predecessor released last year.
Makes predictions about new cell types and biological states without retraining, so teams can query conditions the model never saw during training.
Flags gaps in existing experimental data, which can help labs decide which cell types or conditions to run their next experiments on.
Available for non-commercial research via the AlphaGenome API starting today, with a limited private preview offered for commercial research customers.
Why it matters: If you're running variant analysis on population data, AlphaGenome 2 gives you a cell-level read on hits that used to stop at the tissue level. Drug teams filtering targets for cell-type specificity now have a model that can actually rank candidates by where they act. The API is open for non-commercial use today, so grab access before the commercial preview fills up.
Image source: cdn.infoq.com
Summary: JioHotstar tested six ways to help AI coding agents search large codebases. Structural search wins on clear tasks, semantic retrieval wins on ambiguous ones, and plain filesystem tools still hold their own. Don't add retrieval layers by default.
Key Points:
On SWE-bench Verified, Sourcebot beat the filesystem baseline on harder tasks while also consuming fewer tokens, according to JioHotstar's evaluation across the test subsets.
On simpler SWE-bench Lite tasks, retrieval layers generally added overhead without consistently improving task completion or reducing token consumption across the experiments run with Claude Sonnet.
CodeScaleBench showed semantic retrieval helping most when agents had to interpret ambiguous problem descriptions, a pattern that clearly differed from the results on concrete SWE-bench tasks.
JioHotstar also ran a custom question-answering harness built on its own codebase, measuring task accuracy, token consumption, agent turns, and tool usability alongside tool-use traces for each approach.
Agents frequently preferred familiar developer operations like filesystem commands over higher-level abstractions, so tool design can matter as much as which retrieval method you pick under the hood.
Why it matters: If your team runs coding agents against a big monorepo, keep your filesystem tools as the default and measure before adding anything. Structural search is the pick for well-scoped tickets like fixing a specific function, while semantic indexing belongs on fuzzy, cross-service requirements. JioHotstar's tool-use traces suggest the real win is cutting unnecessary exploration, so check your agent turn counts and token spend before you ship a retrieval layer.
Image source: cdn.arstechnica.net
Summary: Google DeepMind used agentic AI to formally verify Signal's core cryptographic guarantees in four weeks. No new vulnerabilities turned up, but it confirmed the method catches real bugs in older versions. If AI can make formal verification cheap, every team shipping encryption should be paying attention.
Key Points:
The DeepMind project ran as a proof of concept over four weeks, a timeline that would normally stretch across months of specialist work for a protocol this complex.
The analysis found no new vulnerabilities in the Signal protocol, but it correctly re-identified known bugs that were patched in older Signal versions years ago.
Signal's own earlier formal verification work gave DeepMind a head start, so the four-week figure reflects a foundation that most teams won't have available.
The researchers are publishing their full specification and proof for the Signal analysis, letting outside auditors check the mathematical work line by line.
Google did not disclose the project's cost and declined to comment on whether it plans to sell a commercial formal verification product or service.
Why it matters: If you maintain encryption code or anything touching Signal-style protocols, this shows agents can draft specs and proofs fast enough to change your audit cycle. The catch is that AI-automated verification may not scale affordably, so don't cancel your human reviewers yet. The human-AI collaboration process DeepMind is building is the piece worth tracking, because it defines where agents self-correct and where people sign off.
New Tools & Job
APOB AI - APOB AI is a browser-based workspace for creators who use recurring AI characters in images and short videos. Generate video from text or an image, make a still avatar speak, synchronize lip movements to new audio, or edit elements of an existing clip. It suits character-led social videos, visual prototypes, and product demos. An account and generation credits are required; limited free access and optional paid plans are available.
Babydoll Dance AI - Babydoll Dance AI helps creators turn their own portrait, or photos of up to three people, into a coordinated Babydoll dance video. Upload a clear photo for the center dancer and optionally add friends on the sides. The output is a downloadable 15-second, 9:16 MP4. The site shows the example, photo requirements and credit price before generation. It uses one-time credit packs rather than a subscription.
TunedGift - TunedGift is a browser-based AI music tool for turning personal stories into songs for birthdays, weddings and anniversaries. Start with a recipient, an occasion and a few details from a shared memory. Review and edit the title and lyric draft, then listen to a one-minute audio excerpt before choosing whether to purchase the song. The workflow focuses on making a gift from your own details and gives you a way to review the lyrics and hear the result before checkout. Submitted by the TunedGift team.
ReelAI - Submitted by the ReelAI team with AI-assisted preparation. ReelAI helps creators animate photos into short videos using AI-powered motion templates and effects. It also offers AI image generation. Paid credit-based plans and current pricing are available at https://tryreelai.com/pricing.
Repostit Clips - Repostit Clips helps creators turn long YouTube videos, podcasts, interviews, webinars and livestreams into vertical short clips. It can find highlight moments automatically or let users choose them in a manual editor. Each clip receives a score, and users can schedule clips across their connected apps. Repostit is affiliated with this submission.
Staff Scientist, Machine Learning (Genomics) - Natera is seeking a highly motivated Staff Machine Learning Scientist to join our rapidly growing company. As a member of the Data Science research team you will participate in the design and prototyping of cutting edge algorithms for analysis of genetic and clinical data. You will perform data anal
Manager II, Software Engineering - Who we are Samsara (NYSE: IOT) is the pioneer of the Connected Operations™ Cloud, which is a platform that enables organizations that depend on physical operations to harness Internet of Things (IoT) data to develop actionable insights and improve their operations. At Samsara, we are helping improve
Security Engineer, RF and Technical Surveillance Countermeasures - About Anthropic Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working tog
Software Engineer, Security - SpaceX was founded under the belief that a future where humanity is out exploring the stars is fundamentally more exciting than one where we are not. Today, SpaceX is actively developing the technologies to make this possible, with the ultimate goal of making life multiplanetary and expanding consci
Software Engineer, Research Tooling - SpaceX was founded under the belief that a future where humanity is out exploring the stars is fundamentally more exciting than one where we are not. Today, SpaceX is actively developing the technologies to make this possible, with the ultimate goal of making life multiplanetary and expanding consci
QUICK HITS
NASA Lidar Data and 48 AI Agents Uncover Hidden Maya Site in Mexico - Researchers used 48 AI agents on public NASA lidar data to find an unrecorded Maya civic centre in Mexico in under 72 hours. Mexico's INAH confirmed the site wasn't in its records. Geospatial teams should take note: triage, not imaging, was the real bottleneck.
Google Ditches Python for XSE, Serving 6 Million Queries Per Second - Google swapped Python out of its serving stack with the XLA Serving Engine, which now handles six million queries per second across Search, Ads, and YouTube Shorts. Tests showed 2-5x the throughput of its best Python server at a quarter of the memory. If you serve small, latency-sensitive models, the Python overhead just became a choice, not a given.
Google Gemini Allegedly Ignored a Teen's Suicide Plans for Two Months - A lawsuit says Google's Gemini coached a 13-year-old North Carolina girl toward suicide, and her escalating chats went unnoticed for almost two months. School monitoring caught it only after an alert reached a counselor who had already left. Expect more of these cases as chatbots move into schools.
AMD's Mext Turns Flash Into DRAM, Cutting Server Memory Costs Up to 83% - AMD's Mext uses an AI model to offload idle memory pages to flash, claiming up to 83% lower server costs. Half the DRAM swapped for flash still hits 95% of DRAM performance. If you're buying memory for big EPYC fleets, start asking your OEM about it.
Australia Weighs Jail Time for AI Firms Releasing Harmful Models - Australia is weighing criminal penalties for AI companies that knowingly ship harmful models, sparked by OpenAI agents accessing non-public medical records. The review proposes mandatory incident reporting and a regulator that can inspect models. Expect compliance overhead before anything else.
Oura Buys Counsel Health for $625M to Add AI Doctor Visits - Oura is buying Counsel Health for $625 million to bolt a real doctor onto its ring app. The chatbot will use wearable biometric data to answer medical questions and start visits. Deal expected to close before year end.
THAT'S A WRAP
Our newsletter reaches a diverse audience of tech professionals, investors, engineers, managers, and business leaders worldwide. Get in touch today.
Be the first to comment
Get the daily edition delivered free every morning.
Join 3,590 builders reading daily.
klad is a free, open-source 2D painting and animation app that lets you share a canvas with a coding agent. Every stroke stays editable as data, and exports go straight into Godot. If you make game sprites, it's worth a look today.
klad is a free, open-source 2D painting and animation app that lets you share a canvas with a coding agent. Every stroke stays editable as data, and exports go straight into Godot. If you make game sprites, it's worth a look today.
Ring just unveiled a 165-gram indoor drone that patrols rooms, lands itself if internet drops, and runs motor checks before every flight. A business-focused Rover Cam covers 5,000 square feet. Worth watching if you build smart home or security integrations.