[AINews] Muse Spark 1.3 matches GPT-5.6-Sol, confirming Meta Superintelligence as the newest Frontier Lab, >90% discount for training

Launch season continues from yesterday , with Gemini 3.8 Flash as rumored today, but Muse Spark 1.3, promised in Zuck’s big comeback letter last month, definitely deserved the title story win today. Per AAII it is now the #3 model in the world (!?!) Just look at the confidence displayed finally putting up comparable numbers to the frontier models from OpenAI and Anthropic (Opus, not Fable)… and promising that it will be open weights as well(!!!): They have an interesting pricing model where it is 90%+ cheaper if you opt in to training: AI News for 8/22/2026-8/24/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space . You can opt in/out of email frequencies! AI Twitter Recap Agent Engineering Courses, Curricula, and Developer Practice Stanford is formalizing AI-native software engineering as a discipline : @mihail_eric announced a new edition of The Modern Software Developer centered on what he calls the “2026 metamorphosis” of software engineering. The notable signal is not just the course itself, but the curriculum reset: 85% of Fall 2025 material is being replaced with topics like agent skills, context engineering, MCP portals, agent-ready codebase design, agentic code review, security, parallel background agents, and software factories . The course also requires students to ship PRs into real OSS repos with support from partners including Browserbase, OpenHands, Semgrep, Milvus, Marimo, CrewAI, Warp, Vercel, Unsloth, and Anyscale, among others. A second Stanford course focuses on first-principles agent construction : @Diyi_Yang and @michaelryan207 announced CS329Z: Engineering AI Agents , explicitly framed around building agents “from scratch.” Alongside Mihail Eric’s course, this suggests a broader shift from “prompting” pedagogy to systems-oriented agent engineering : harnesses, evaluation, memory, tooling, orchestration, and production constraints rather than model usage alone. Practitioner discussion is converging on stateful intelligence allocation, not simple routing : In a panel prompt, @HarryStebbings highlighted @EnoReyes’s argument that getting the most out of models requires more than routing—agents need to understand task state, what just happened, and what comes next in order to allocate intelligence dynamically. That lines up with @jerryjliu0 ’s point that vendor-neutral startups can outperform frontier labs on narrow tasks by optimizing the harness end-to-end and selectively using both frontier and open-weight models. Model Architecture and Inference: Astra Rumors, Looped Transformers, and Real-Time Serving The “Astra is a looped transformer” rumor is probably less novel than headlines suggest : @rasbt unpacked reporting around OpenAI’s rumored Astra architecture and argued that the cited “recurrent depth” or “looped transformer” concept is a fairly modest architectural tweak rather than a breakthrough on its own. He points to Nanbeige 4.2-3B as an open-weight precedent: a 22-layer transformer stack reused twice , effectively behaving like a 44-layer model without doubling parameter storage. The tradeoff is straightforward: similar memory footprint, roughly ~2x compute , and only partial token-efficiency retention versus a standard stack. The more substantive historical reference is Mixture-of-recursions , where a learned router adaptively determines how many passes a token gets, allowing easy tokens to exit early and hard tokens to receive more compute. Hidden reasoning is not a necessary implication of recurrence : A second important clarification from @rasbt is that layer reuse does not inherently “obscure chain-of-thought” . It simply moves more computation into latent activations before token emission. If recurrent depth reduces visible reasoning traces, that’s because the model may need to emit fewer intermediate tokens, not because looped transformers intrinsically suppress textual CoT. Serving infra updates continue to target realtime multimodal workloads : @vikhyatk announced Photon 2.1 , adding text-to-speech models and NVIDIA B200 support to a realtime multimodal inference engine. Separately, Baseten announced hosted availability of GLM-5.3 Fast , emphasizing higher TPS and real-time deployment positioning via @baseten . Agent Harnesses, Skill Retrieval, and RL Post-Training Tooling ByteDance Seed’s HarnessDev reframes agent evaluation around the harness, not just task completion : @omarsar0 highlighted a new paper on HarnessDev , which asks models to start from a weak but runnable seed and build an execution harness, then improve it in a second stage using downstream feedback. Both stages are scored on capability and execution-token cost , making efficiency part of the objective. Across six creator LLMs, four domains, and 2,207 held-out downstream instances , generated harnesses still lag mature human-engineered systems on code, search, and research , but match or exceed them on writing and ML experimentation . The key nuance is that self-evolving harnesses help, but gains are unstable, model-dependent, and only partially transferable . Related ecosystem signal: exo and recursive self-improvement tooling : @omarsar0 also called out the exo harness as a useful entry point for understanding recursive self-improvement workflows, indicating a growing interest in frameworks where agents improve not just outputs but their own scaffolding. Skill retrieval may look good in aggregate while hurting the tasks that actually trigger it : @dair_ai summarized a paper proposing Retrieval-Invoked Actual-Use Effect , a matched-evaluation method that runs the same task twice , with and without skills enabled, and only counts tasks where retrieval actually fired. Across 17 LLMs on coding and math, the paper finds cases where retrieval improves overall scores while having a negative same-task effect on the subset of tasks where it was used. For teams maintaining skill libraries or tool directories, this is a practical warning against over-interpreting aggregate lift. RL post-training infra is becoming more productized : The SGLang team promoted an event with Baseten and NVIDIA Dynamo around Miles , an RL training framework that uses SGLang as the rollout inference engine for faster, more reliable RL post-training @sgl_project . @AravSrinivas separately described Miles as open-source RL-as-a-service , reinforcing the trend toward reusable post-training stacks rather than bespoke internal pipelines. Google Gemini 3.8 Flash Cyber and Production Friction Around Google Tooling Google introduced a specialized cybersecurity model with strong benchmark claims : @sundarpichai announced Gemini 3.8 Flash Cyber , positioned as Google’s most capable cybersecurity model while retaining Flash-level speed and pricing . Reported numbers include 86.2% on CyberGym , 47.2% on CWE-Bench for patching , and 70%+ success on an internal vulnerability-discovery benchmark across 20 programming languages . At the same time, developer sentiment points to harness and account-risk concerns : @theo argued that Google currently has weak developer ergonomics around harnesses, code apps, third-party integration, and especially aggressive bans tied to core Google accounts . @QuinnyPig sharpened that concern, noting the blast radius can extend beyond Gmail/Workspace to Google Cloud accounts associated with the same identity . Theo’s later complaints about slow, tool-call-heavy coding behavior on Gemini tasks ( 1 , 2 , 3 ) are anecdotal, but they underline the gap between benchmark performance and production developer UX . Meta Muse Spark 1.3 and the Video/Multimodal Release Cycle Meta launched Muse Spark 1.3 for agentic and coding workloads : @shengjia_zhao introduced Muse Spark 1.3 as the strongest model