It’s an odd thing to say “not much happened” while running AIEWF workshops, but objectively, that is true - vibes were good but the wider world collectively took a breather to process that shock Germany loss today. In the meantime you can think though how to build better Skills, which is emerging as a top theme of the conference throughout the week.

and help us turn notifications on for the first keynote in 9 hours:

AI News for 6/27/2026-6/29/2026. 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!

• Meta’s non-invasive brain-to-text milestone drew the biggest technical attention. @AIatMeta announced Brain2Qwerty v2, a real-time sentence decoder from raw brain signals; @JeanRemiKing summarized the release and links; @AIatMeta added that Meta is releasing the training code for v1/v2 and BCBL is releasing the v1 dataset.

• Cursor shipped iOS + remote agents in one of the day’s biggest product launches: @cursor_ai introduced Cursor for iOS with always-on cloud agents and remote control of agents on your computer; follow-up tweets highlighted Live Activities and diff review on phone.

• Open-weight model access is being productized, not just discussed: @cline launched a $9.99/mo pass for discounted access to GLM 5.2, DeepSeek, Kimi, MiniMax, Qwen, etc.; @cognition introduced Devin Fusion, claiming 35% lower cost for “Fable-level” coding via a hybrid-model harness.

• Arena crossed meaningful commercial scale: @arena and @ml_angelopoulos said Arena reached $100M ARR run rate eight months after launching its evaluation product, with a platform now emphasizing post-deployment and agent evaluation.

• Infrastructure pressure remains a first-order theme: @kimmonismus argued China’s energy, data center, and domestic-hardware strategy is becoming a serious strategic threat; @garrytan condensed the operational response to “Build power and datacenters.”

Brain-computer interfaces and AI-for-science tooling

• Brain2Qwerty v2 is the clearest research release of the day. Meta says the system decodes words and semantics, not just characters, from non-invasive recordings in real time, narrowing the gap with invasive BCIs. Community summaries highlighted reported jumps from prior non-invasive results to ~61% word accuracy overall and 78% for the best participant, trained on data from 9 volunteers in controlled typing settings. The key engineering point is not consumer readiness, but that the stack combines raw neural-signal modeling with language modeling strongly enough to make sentence-level decoding practical in the lab. See Meta’s announcement, the code/data release details, @JeanRemiKing’s thread, and a cautious external summary from @kimmonismus.

• The release also became a datapoint for agent-assisted research. @stalkermustang pointed to Meta’s note that an Auto Research workflow, powered by a coding agent, discovered and implemented improvements that reduced word error rate beyond standard HPO. Whether or not one buys the “vibe-science” framing, the more sober takeaway is that coding agents are increasingly useful for closed-loop experimental iteration on ML systems, not just repo scaffolding.

Inference systems: DSpark, vLLM, and decoding mechanics

• DeepSeek’s DSpark was the most substantive inference topic. A long explainer from @ZhihuFrontier framed DSpark as an important step in speculative decoding, with emphasis on two ideas: better draft generation and smarter verification scheduling. Reported gains include 30.9% higher accepted length vs Eagle3 and 16.3% vs DFlash on Qwen3-4B, plus production deployment in preview engines for DeepSeek-V4-Flash and V4-Pro. Follow-on commentary from @teortaxesTex and @vllm_project underscored the practical consequence: DSpark looks like a new SoTA single-GPU spec decode path, and the vLLM community is already integrating it.

• More broadly, several tweets sharpened the mental model of current inference bottlenecks. @_avichawla gave a solid explainer of prefill vs decode, TTFT vs inter-token latency, and why decode is often memory-bound because of KV-cache reads. This is useful context for why speculative decoding, KV-cache optimization, grouped-query attention, and attention redesigns matter more than raw FLOPs in many production workloads.

• NVIDIA/vLLM also pushed practical self-hosting: @vllm_project highlighted a guide for serving Nemotron-3-Ultra 550B with four DGX Spark boxes behind a single OpenAI-compatible endpoint. The notable part is less the stunt than the normalization of private, multi-node frontier-ish inference using standard serving stacks.

Agent harnesses, routing, and multi-model orchestration

• The center of gravity in agent systems continues to move from “pick the best model” to harness engineering. @cognition launched Devin Fusion, a hybrid-model coding harness claiming 35% cost reduction while maintaining “Fable-level” quality. @walden_yan described related work around sidekick and mid-session routing, and @jerryjliu0 noted the cache-efficiency advantage of sidekick-style delegation. The emerging pattern: keep an expensive planner in the loop, hand bounded subtasks to cheaper models, and preserve cache locality/context continuity.

• Dynamic subagents became another common motif. @LangChain, @sydneyrunkle, and @hwchase17 all highlighted workflows where the main agent writes orchestration code rather than merely invoking tool calls. This is notable because it shifts the abstraction from “tool-using chatbot” to something closer to a programmable control plane for large task fanout.

• Open routing and retrieval stacks also got more concrete. @LlamaIndex and @jerryjliu0 introduced a Retrieval Harness combining semantic search, grep, file listing, and file reading in one agent loop—essentially a rebuttal to simplistic “grep is all you need” positions also criticized by @max_paperclips. On the eval side, @hwchase17 announced a Trace Judge model for detecting trajectory errors at ~1/100th the cost of closed models.

Open models, Chinese labs, and commercialization of access

• GLM 5.2 remained the focal open model in discussion, not because of an official launch today but because many builders are now treating it as a default serious option. @cline productized access with a monthly pass bundling GLM 5.2, DeepSeek, Kimi, MiniMax, Mimo, and Qwen, reducing friction around API keys and provider churn. @tonbistudio tested Mixture-of-Agents configurations using GLM 5.2 with Kimi and MiniMax. @Astrodevil_ used GLM 5.2 as the driver for a DevRel content-research agent.

• A second thread is the continued acceleration of Chinese open-weight competition. @eliebakouch flagged an upcoming LongCat 2.0 / Owl Alpha model from Meituan: 1.6T total / ~48B active, 1M context, 35T training tokens, n-gram embeddings, sparse attention, and training on 50k Chinese accelerators. @sun_hanchi framed this as potentially the first near-frontier model trained at this scale on domestic Chinese hardware. Even allowing for uncertainty in the hardware details, this is strategically meaningful.

• On the policy/commercial side, open-source proponents argued that clampdowns on frontier APIs may backfire by pushing developers toward weights they control. See @theinformation, @ClementDelangue, and @MTSlive for the recurring theme that open weights are structurally harder to suppress than APIs.

RL, training infrastructure, and benchmark/eval platforms

• Snowflake Arctic RL is one of the stronger infra releases in the batch. @StasBekman announced an open-source project integrating with VeRL and SkyRL, featuring ZoRRo for up to 6x actor-update acceleration and 3.5x end-to-end speedup, reducing a Text2SQL training run from roughly 5 days to ~36 hours on 32 H200s. Snowflake also claims its Arctic-Text2SQL-R2 beat tested configurations of Gemini 3.1 Pro and Claude 4.7 on its enterprise SQL benchmark, with open recipes for text-to-SQL and multi-hop QA.

• Arena continued its transition from benchmark project to evaluation company. @arena and @ml_angelopoulos reported 700M+ conversations, 82M+ votes, and over 10M monthly visitors, with newer emphasis on agent-mode evaluations like task completion and hallucination rates. That makes Arena increasingly relevant as a post-deployment CI/CD layer for models, not just a preference leaderboard.

• Several other releases fit the same trend toward specialized infrastructure: @wandb launched ARIA, an autoresearch agent inside W&B; @agenticin promoted Micro-Agent routing; and @fitsumreda introduced Nemotron-TwoTower, which clones an AR LLM into a diffusion-style parallel generator, claiming 98.7% AR quality at 2.42× throughput for a 30B model.

Platform and developer product updates

• Cursor’s mobile/remote push is notable because it makes “cloud agents from your phone” feel operational rather than aspirational. The product now supports launching always-on cloud agents and remotely controlling computer-bound agents from iOS, with PR diff review and notifications in-app (launch, details).

• Claude on Azure Foundry is now GA. @Azure, @claudeai, and @ClaudeDevs said customers can run Claude Opus 4.8 and Haiku 4.5 in Microsoft Foundry with Azure identity, billing, governance controls, prompt caching, and thinking support.

• Rampart from @ndstudio stood out as a pragmatic privacy tool: a 14.7MB browser-side model for redacting PII before data leaves the client. For teams trying to make AI usable in regulated settings, this kind of small, local preprocessing model may matter more than another general-purpose chat UI tweak.

• GLM-5.2 753B (IQ1_S) fully local across 2×M5 Max over one TB5 cable — ~16 tok/s, llama.cpp RPC [video] (Activity: 377): A user reports running GLM-5.2 753B fully locally using Unsloth dynamic IQ1_S quantization: nominally ~1.6 bits but ~2.1 effective bits due to mixed higher-precision layers, yielding a 202GB on-disk model. The setup shards weights across 2× M5 Max systems with 128GB unified memory each over a single Thunderbolt 5 link using llama.cpp RPC, keeping all weights resident with no SSD paging and achieving ~16 tok/s generation, 16k context, and q8 KV cache; TTFT is prompt-length dependent due to prefill. Commenters found 16 tok/s for a 753B model over two Macs surprisingly high, with one asking whether the video appeared faster than reported. Another noted the setup is impressive but questioned how the very low-bit 753B quant compares on complex reasoning against a smaller higher-precision model such as a 70B at 4-bit.

  • A commenter questioned whether the reported ~16 tok/s for GLM-5.2 753B IQ1_S across 2× M5 Max over Thunderbolt 5 was accurate, noting the video appeared faster; another highlighted that while the throughput is impressive for a 753B local setup, the very low-bit IQ1_S quantization raises the technical question of reasoning quality versus a smaller 70B at 4-bit model.

  • One user provided comparative llama.cpp RPC-style benchmarks using an M3 Ultra Studio 256GB + M3 Max MBP 128GB running GLM-5.2-UD-IQ4_XS: 13.03 tok/s at 2,377 context tokens with TTFT 3.09s, 8.64 tok/s at 22,485 context with TTFT 2.33s, and 6.21 tok/s at 32,595 context with TTFT 5.53s. They clarified that TTFT included cache prefill, making the measurements more comparable for long-context generation.

  • Another commenter asked whether multi-Mac connectivity is already supported in llama.cpp or requires a custom driver, pointing to the implementation-level question around whether this setup uses built-in llama.cpp RPC capabilities or bespoke Thunderbolt networking/inference orchestration.