AI Frontier Daily · October 25, 2026
AI Frontier Daily · October 25, 2026
Daily global AI trends digest, caught up in five minutes.
Headlines
1. Block’s Dessa Sounds Eerily Human: A Full-Duplex Voice Model Trained on Speech Alone, Running Locally
On October 23, Jack Dorsey’s Block released Dessa from its research lab — billed as “the most authentic full-duplex speech model ever built.” It can talk and listen simultaneously, like OpenAI’s GPT-Live, complete with laughs, pauses, and cries. Unlike mainstream speech stacks, Dessa was trained entirely on human speech data (on a single NVIDIA GPU rack), derives its knowledge from conversational data rather than a bolted-on LLM, and can run locally on consumer GPUs (demonstrated on a MacBook M4 Max). The team calls it “a GPT-3 moment for speech,” and says Square’s existing voice-ordering use cases can move on-device at lower cost.
Source: TechCrunch / Block
2. Two Weeks Before the Midterms: Forced to Pick, Grok 4.7 Chose Republicans 92% of the Time, GPT-6.1 Sol Democrats 97%
A study by Convovo, a conversational-AI research studio led by Cornell professor Tapan Parikh, forced models via API across roughly 500 races to “name one candidate, don’t hedge”: SpaceXAI’s Grok 4.7 recommended the Republican in 92% of answers, OpenAI’s GPT-6.1 Sol the Democrat in 97% — the two agreed less than 10% of the time. Consumer apps refused to weigh in (OpenAI notes the API doesn’t reflect the ChatGPT experience), but the researchers say hidden leanings seep into everyday questions. Tracing the sources: Grok frequently surfaced conservative outlets — a Christian conservative voter guide in 39% of answers, Turning Point Action 23%, X 41% — while GPT leaned on Wikipedia, the AP, Axios, and Reddit (OpenAI has content deals with the latter three). Each Grok API request also carried roughly 1,240 tokens of input the researchers didn’t write and couldn’t remove — potentially “hidden instructions.” With internet search off, GPT’s Democratic lean persisted (Democrat in 24 of 30 races — it’s in the weights), while Grok’s Republican lean vanished (it refused 27 of 30, and all three picks it did make were Democrats — for Grok, it’s the search path).
Source: TechCrunch / Convovo / Pew
3. Tsinghua AIR’s “Agent Hospital”: Multi-Agent Doctor Collaboration Puts an Entire Specialist Team Behind Every Physician
Zijing Zhikang, incubated by Tsinghua AIR (founded by academician Zhang Yaqin), has showcased its multidisciplinary human-agent collaboration technology and upgraded its “Agent Hospital” system, already in pilot across hospitals nationwide. Each doctor gets a dedicated AI doppelganger: for same-department cases, a single agent assists the clinic; for cross-specialty complexity, agents with different expertise convene a “multidisciplinary consultation,” pushing cross-specialty support earlier into first visits and pre-consultation. The system deliberately does not smooth over disagreement — divergent agent opinions are highlighted side by side, and the human doctor makes the final call using patient context and evidence. The doppelgangers keep learning from their human doctors’ decisions (self-evolution), with decision authority always returning to the physician. The pitch: instead of “AI replacing doctors,” every patient walks into the clinic with an on-call expert team.
Source: QbitAI / Zijing Zhikang
Models & Platforms
4. Chinese 5B Image Model MOSS-Image Goes Viral Abroad: 1.3M Downloads in 4 Days, Trained Without AI-Generated Images
MOSS-Image, an open-source 5B-parameter DiT image model from Mosi Intelligence and the OpenMOSS team, racked up ~1.3 million downloads across Hugging Face and the ComfyUI community in just 4 days, plus 1.4k GitHub stars and the top spot on Hugging Face’s text-to-image trending chart. Western developers are obsessed with how “un-AI” it looks — the training data deliberately excluded AI-generated images, giving outputs a more natural feel. The community quickly shipped 8-step/4-step accelerated versions, INT8 quant, GGUF conversions, LoRAs, and new ComfyUI workflows, and it’s being compared head-to-head with FLUX.3-klein, Krea2, and Qwen Image 2.1. Engineering focus: dense Chinese/English text rendering (96.3% Mean-6 text coverage on LLTB, best among open-weight models) and precise layout control via JSON Layout and bounding boxes — built for posters and e-commerce design.
Source: QbitAI / Hugging Face
5. Recursive Ships Flint, a Laptop Coding Model Largely Built by Its Own AI Research System
Recursive — founded by ex-Salesforce chief scientist Richard Socher with $650M from GV, Nvidia, and AMD Ventures — released its first model, Flint, on October 21: a free coding model that runs on a laptop, built largely by the company’s automated research system, the “Eureka Machine,” which helped at nearly every step from pretraining to post-training to inference, letting a small team build it from scratch in under four months. Flint is a mixture-of-experts model (40B parameters, 2.5B active) that hit ~131 tokens/sec on a MacBook Pro M4 Pro — roughly double Gemma 4 and Nemotron 3.5 Lightning on the same machine — and over 150 tokens/sec on Nvidia’s RTX Spark. It outperformed every sub-50B model on SWE-Bench Verified, with only Alibaba’s Qwen3.8-27B scoring higher on Recursive’s own six-test average. Eureka’s contributions include a sparser model shape, low-bit training that shrank the model 71% with almost no quality loss, and custom kernels. Flint is Apache 2.0 on Hugging Face — an early public test of recursive self-improvement.
Source: TechCrunch / Recursive
6. Dawa Tech’s Scenes AI Wins an Nvidia Award: A Real-Time 34B World Model Running on a Single RTX 5090
At the 2026 Nvidia Startup Showcase semi-final in Suzhou on October 15, Chongqing’s Dawa Tech took the “Outstanding Semi-Final Enterprise” award with Scenes AI, its real-time world model (580+ applicants, 30 finalists). Scenes AI rejects the “predict the next frame” paradigm: instead it maintains a continuously running world state — scene, characters, camera, lighting, objects, and physics are individually readable/writable variables. When something changes, only the changed state (the delta ΔS) is updated; the rest stays stable and queryable. Crucially, its input can be a plain camera feed of a real room: a person is instantly transported into a hand-drawn summer meadow, a raised hand makes it snow indoors, yet the writing on the wall and the chair’s position never drift. The whole system runs locally on a single RTX GPU. As Dawa puts it: “A video is an output. A world is a process.” Architecture: world memory + neural physics + sparse diffusion + neural rendering + AI lighting.
Source: QbitAI / Nvidia
Industry
7. Meta Q3: A New Non-Voting Stock Class to Fund the AI Buildout, and Muse Passes 10M Users in Six Weeks
Meta’s Q3 report (October 21) showed $64 billion in revenue (up 26% YoY), with costs and expenses up 68%. CFO Susan Li confirmed the company will “create a new class of non-voting capital stock,” distributed to existing shareholders via a stock dividend, for capital raising, employee compensation, and acquisitions — while preserving Zuckerberg’s founder-led structure. On AI: “In just a month and a half, more than 10 million people in the U.S. and Canada have set up Muse personal agents,” Zuckerberg said, adding that the app has been number one in the App Store almost every day and shopping is a top activity, with ambitions to eventually “serve billions of people.” Reality Labs losses are expected to shrink to $18–19 billion from 2025 levels.
Source: TechCrunch / Meta
8. Uber Puts All 30,000 Employees on ChatGPT Work: From a 1,000-Person Pilot to Company-Wide
Uber and OpenAI expanded their partnership on October 22: the ride-share giant’s entire global workforce — about 30,000 people — will use ChatGPT Work, the enterprise agent OpenAI launched in July. Uber piloted it with 1,000 employees earlier; roughly 80% of early adopters were non-engineers, using it to analyze operations data, localize training materials for different markets, and build personalized dashboards. Uber’s VP of engineering described “AI pods” — 30 AI engineers paired with teams for two weeks to automate key job functions — as proof of the upside. Uber declined to share spend, saying it will evaluate usage outcomes to guide wider AI deployment. Another marker in the OpenAI-vs-Anthropic enterprise race.
Source: TechCrunch / Uber
9. 1Password for Agents: AI Agents Can Log In on Your Behalf — Without Ever Seeing Your Password
On October 22, password manager 1Password launched “1Password for Agents”: users authorize an agent to log into their personal accounts and get things done, while passwords, secrets, and credentials never leave the vault. Each login attempt exchanges a unique digital key that confirms the agent’s identity, pings the user for approval, and only then does 1Password sign the agent in by filling credentials on its behalf. First wave of supported agents: Anthropic’s Claude Duet, OpenAI’s Dots and ChatGPT Work, Grok Bot, Meta Muse, and Town. The same day, Axios reported that U.S. banks have begun detecting agentic activity inside customer accounts — consumers are letting automated services act on financial matters, yet banks mostly can’t tell whether that activity was user-authorized. The trust layer for agentic commerce has opened.
Source: TechCrunch / 1Password / Axios
Research
10. This Week’s Papers: When Bored AI Agents Play, Confident Falsehoods Erode Trust, Causal CoT, and Instrumental Convergence Under RLVR
From MIT and the October 22 arXiv wave:
AI agents, left alone, started playing (MIT): Agents exploring a virtual island with no specific task began inventing challenges of their own — behavior MIT researchers say bears the hallmarks of play. The report’s closing line: “Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.” A different answer to where capabilities come from: not everything is goal-driven training.
Who Is Sycophantic to Whom? (arXiv:2610.31579): Past work asks whether models flatter users; this paper asks the reverse — when an LLM is confident, consistent, and wrong, do users bend toward it? In a within-subjects experiment, 32 participants learned science topics from five assistants built on the same model (GPT-5-mini), each with 0/25/50/75/100% of facts swapped for plausible falsehoods, all speaking with full confidence. Quiz scores fell from 7.4 to 2.9 out of 10, and the planted falsehood was chosen on 60% of falsified items versus 11% of correctly taught ones (odds ratio 12.5). Confident AI falsehood systematically collapses users’ trust calibration.
Causal CoT (arXiv:2610.31639): Standard chain-of-thought rationales may be post-hoc — dense attention lets a model look up the answer from context and narrate a justification afterward. CausalCoT decomposes problems into a DAG of atomic sub-questions, evaluating each node in an isolated context containing only its direct causal parents, removing attention pathways to non-causal content. Downstream predictions become genuinely sensitive to local reasoning, with reasoning performance preserved — “causal transparency” for CoT.
Questioning Instrumental Convergence in RLVR (arXiv:2610.31624): Instrumental convergence — self-preservation, resource acquisition, deception, resistance to oversight — is often invoked as a general reason to expect power-seeking in capable agents. This work evaluates OLMo-3-7B-Think along its SFT → DPO → RLVR checkpoint trajectory with InstrumentalEval, reframing “does RL make models dangerous” as “it depends on the training regime, the evaluation frame, and the stage of post-training.”
Low-Bit N-Gram Embeddings (arXiv:2610.31646): Many open-weight models use n-gram embedding tables for extra performance — DeepSeek-V4.1-Flash alone allocates 196 billion parameters to its Engram n-gram memory. This paper systematically tests quantization methods and bit allocations: 4-bit tables save roughly 75% of table memory while preserving more than 99% of the validation-loss improvement. A practical answer to how expensive memory tables really need to be.
Source: MIT / arXiv
Compiled from multiple media reports.






