CRUNCHY TECH
YESTERDAY, STILL WORTH KNOWING
Epic bets a delayed Disney shooter on AI-first development while Microsoft and Intel push new silicon, Meta ships rival models and Steam sets another record.
The day's most consequential story is Bloomberg's report that Epic Games has told staff on its unannounced Marvel and Star Wars crossover shooter to adopt an AI-first production model after the project slipped past its expected March 2027 date, partly because Epic underestimated staffing needs. An internal message reportedly declared there would be no more hand-written code, with art generated through full AI workflows using tools such as Anthropic's Claude, and humans managing the output. Some developers were reportedly stunned, joking about asking Claude to fix its own mistakes.
Epic has effectively confirmed the pivot, with CEO Tim Sweeney insisting the goal is not doing the same work with fewer people, though the move lands less than a year after the company laid off more than 1,000 staff. The shooter stems from Disney's $1.5 billion investment in Epic in 2024 and extends a strategy Epic had already signalled with Unreal Engine 6's planned AI integrations. It makes Epic one of the first major studios to attempt a triple-A game built primarily with AI, in a field where plagiarism and output quality remain live concerns.
Microsoft announced Maia 300, its third-generation in-house AI chip, arriving in 2027 and positioned as the primary inference system for Azure, reducing reliance on Nvidia and AMD GPUs. Scott Guthrie claimed 40 percent more tokens per megawatt than leading GPU systems, though those are vendor figures. The striking detail is cadence: Maia 200 was only announced in January and already runs Copilot workloads. Mustafa Suleyman tied the chip to Microsoft's ambition to run its own models across the full stack and catch up with Google DeepMind, OpenAI and Anthropic.
Intel countered with Crescent Island, now named the D-Series, a 350 watt air-cooled server card built purely for AI inference with up to 480 GB of memory, no display outputs and enough capacity to serve 300 billion parameter models on one card. Pricing and independent benchmarks are still undisclosed. Separately, Intel has reportedly raised processor prices for laptop makers by up to 20 percent on the new Panther Lake generation, blaming AI demand for squeezing manufacturing capacity, with costs likely to reach shoppers once existing stock clears.
Meta launched Muse Wonder and Muse Wonder Full for Muse Code subscribers, claiming benchmark results that rival OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5. Those are Meta's own tests, and how the models stack up against Google's Gemini 4 Argon remains unclear. The launch continues the Superintelligence Labs push that began with Muse Spark.
A Microsoft Research study offers a curious counterpoint to all this model ambition: long-running AI research agents perform better when their chat history is wiped. Accumulated context traps agents in what the researchers call an idea basin, and their Fork-and-Flush method, which branches agents with fresh context but shared workspaces, scored 0.78 against 0.47 for single runs. Even simply clearing context improved results by 22 percent. Meanwhile the market for Jev-style decision models, which return probabilities rather than generated text, has filled with lookalikes weeks after TypeSafe's launch, amid unresolved copying claims from a founder who says he published the architecture a year earlier.
Steam hit 43.5 million concurrent users, its second record this year after January's 42 million, with in-game players topping 14 million. Oddly the peak came just after the Autumn Sale ended, helped by Aion 2's worldwide release and a Deadlock update. Elsewhere, a coder has built an open source tool that turns original PS4 files into a native PC version of P.T., Kojima's long-discontinued playable trailer, adding upscaling, VR support and a museum of unused content.