AI News

An AI agent beside any Mac app. Sider Omni sidebar sees what's on your screen and acts on it directly, right where you're working. Watch the result, ask for the next change. Nothing to copy, nothing to upload. Built on GPT-6 Astra, and it can run on your existing ChatGPT or Codex credits.

AI agents need better computer tools, not just better models. Agent Interface is an open-source layer built to cut repeated screenshots, model calls, and waiting. It reuses learned interactions, runs action-and-feedback loops locally, and asks the model when fresh judgment is needed. Start with the runnable desktop research preview, and follow the Astra/Freedoom experiments exploring control in a world that doesn't pause while AI thinks.

We study length inflation in on-policy distillation (OPD), where student responses can become excessively long and even exhaust the generation budget. We identify \emph{termination-token mismatch} between base students and post-trained teachers as an important source of this behavior. Across Qwen3, Llama, and Gemma, the two models can place their stopping probability on different EOS tokens, even when their declared stopping sets are identical. This mismatch can suppress the student's preferred termination action without reliably transferring the teacher-preferred alternative. We show that aligning the decoding stopping set alone is insufficient, while treating functionally equivalent EOS tokens as a shared semantic stopping action substantially mitigates mismatch-induced length inflation across all three model families. To further understand how termination behavior evolves over training, we study OPD across different K2-Horizon training stages. This stage-wise analysis shows that termination preferences can shift substantially during training, while also revealing a distinct length inflation late in the OPD run that persists beyond termination alignment. Together, these results identify termination mismatch as an important, but not exhaustive, source of OPD length dynamics. We release an implementation incorporating the proposed termination-handling corrections.

Coding harnesses shape how autonomous coding agents translate model capabilities into long-horizon software-engineering performance, yet existing work typically evaluates harnesses as monolithic systems, leaving the effectiveness of individual components unclear. To enable component-level comparisons, we study this question with a lightweight coding harness whose execution loop is fixed while three components are varied: planning, action space, and context management. Across four models evaluated on SWE-Bench Verified and Terminal-Bench 2.1, we evaluate 176 matched settings spanning five context-management strategies, four context-window budgets, and targeted ablations of planning and action space. We find that: (1) Context management becomes increasingly valuable as the context-window budget tightens, with most of its benefit coming from preventing context-overflow failures. (2) Staging rule-based elision before LLM-based summarization provides the strongest overall efficiency among the context-management strategies, whereas making elided content recoverable adds machinery that models rarely use and yields no accuracy gain. (3) Planning shifts from an accuracy scaffold for weaker models to a cost saver for stronger models, with little change in accuracy. (4) Predefined tools improve performance for models with weaker bash proficiency, whereas bash-capable models can operate effectively with a bash-only interface and achieve substantially lower cost, especially on command-line-centric tasks. Trajectory-level analysis explains these effects: context management extends execution trajectories without substantially altering agent behavior, planning changes where trajectories stop, and the action space changes the granularity at which code is written. These findings inform model- and budget-aware harness design and provide a modular framework for evaluating future harness components.

From August 22 to 26, 2026, more than 2,000 humanoid robots from 16 countries gathered in Beijing, China, for the second World Humanoid Robot Games. Over five action-packed days, these metallic athletes competed in 51 events. This was a big increase from the 2025 games. That year, about 500 robots competed in just 26 events.

Robot brain breakthrough could arrive by mid-2027, Spirit AI says, as China’s humanoid robotics sector targets real-world tasks.

An Empirical Study of Harness Design for Coding Agents

paper: https://t.co/pqcxSfs1h0

Startup funding over the past day is being pulled toward a surprisingly consistent set of bottlenecks: AI compute, sovereign chips, power systems, optical interconnects, robotics data, quantum infrastructure, security, and regulated payments. The largest transaction is in a class of its own. Denver-based Crusoe closed the initial $3.9 billion Series F at a $30.9 billion

Manus, which earlier this year had to break off a merger with Meta, is in discussions to raise $500M at a $4B valuation.

Most outbound dies on one decision: who to write to this week. Ami has already made it 17,707 times and kept score. Built on GPT-6 Astra and every one of those campaigns, it reads your website, picks the buyers who actually convert, and plans backwards from the number you need: leads, senders, weeks. If the list you asked for won't work, it says so before it builds anything. Approve the plan and it runs the outreach, watches every campaign, and fixes what slips. 19,854 booked meetings behind it.

Yoetz is a local-first, open-source work ledger for AI coding agents. An agent publishes what it did; Yoetz checks the record and issues a receipt that states what was verified, at what coverage, and what is still open. Supports Codex, Claude and Cursor

Test-time scaling can improve large language model reasoning by generating and combining multiple candidate responses. In sampling-based methods, the inference budget is often described by the number of generated candidates, N. However, N tells us how many candidates are generated, not how they are executed. The same candidate budget can be produced in one batched generation call or split across several sequential calls with smaller batch sizes. We first study the effect of increasing N on reasoning accuracy using Phi-3-mini and Qwen2.5-1.5B on 500 GSM8K prompts. As expected, increasing N from 1 to 8 improves accuracy by 8.4 percentage points for Phi-3-mini and 18.4 points for Qwen2.5-1.5B. However, accuracy alone does not show the systems cost of using a larger candidate budget. We therefore fix N = 8 and compare four generation schedules: 1x8, 2x4, 4x2, and 8x1, where axb denotes a generation calls with b candidates per call. We measure latency, throughput, GPU-hours, and gross GPU-device energy while keeping the total candidate count fixed. On A100 GPUs, eight serial calls use 4.64-4.86x as much gross GPU-device energy and have 5.77-6.12x the P95 latency of one batched call with eight candidates. The same pattern appears across three independently scheduled A100 nodes per model and in short-output SciQ/V100 experiments. These results show that candidate count alone is not enough to describe the systems cost of multi-candidate test-time scaling. When candidates are independent and memory allows it, fewer generation calls with larger batch sizes are more efficient. Evaluations should therefore report not only candidate count and accuracy, but also generation schedule and GPU-level systems metrics.

Large Reasoning Models (LRMs) achieve strong performance on complex tasks but exhibit systematic inefficiency: they often overthink easy problems and underthink hard ones. Existing approaches based on uniform length penalties or rigid routing incur an efficiency tax, trading reduced computation on easy instances for accuracy loss on hard instances. We formulate efficient reasoning as an instance-adaptive computation allocation problem and propose When2Think, a post-training framework for hybrid reasoning that dynamically allocates computation based on problem difficulty. Our method introduces Instance-level Difficulty-Aware Control (IDAC), a reward-shaping mechanism that leverages pre-computed reference statistics (accuracy and token usage) to regulate reasoning depth. Combined with verifier-based rewards and batch-wise standardized advantages, IDAC enables stable critic-free optimization without learned reward models or online reference-model queries. When2Think encourages direct answering on easy instances while preserving extended reasoning on hard instances, thereby learning when to use System 1 (NoThink) versus System 2 (Think). Experiments on mathematical benchmarks demonstrate improved accuracy-efficiency trade-offs: on AIME24, Pass@3 increases by 10.0% while token usage is reduced by 27.9% relative to the base model, and on AIME25, When2Think achieves 40.0% Pass@3, outperforming compression and routing-only baselines.

Been integrating Jev into my custom harness.

I couldn't be more excited about the results I am seeing.

Most demos on my timeline are flashy but very basic. Notice they are all about some boring classification task that was already possible before.

I get it. Jev is fast and cheap. That point was …

UP.Labs, now doing business under the name Vantora, is building startups for industrial corporations.

British AI startup Mantic has raised $25 million in seed funding after its forecasting system beat every human competitor in the summer 2026 Metaculus Cup, giving the young company a striking proof point for its claim that machines are getting better at predicting uncertain real-world events. Radical Ventures led the round at an undisclosed valuation,

GLM-5.3-FlashX is the high-speed variant of Z.ai's GLM-5.3-Flash, a native multimodal model delivering inference speeds of up to 200 tokens/s. $0.37 per million input tokens, $1.25 per million output tokens. 1,048,576 token context window, maximum output of 131,072 tokens.

Providing universal access for developers, the card installs in a desktop, workstation, industrial PC or single-board computer   LAGUNA HILLS, Calif. – Sept. 17, 2026 – BrainChip Holdings Ltd. (ASX: BRN, OTCQX: BRCHF, ADR: BCHPY), a global leader in ultra-low power, fully digital, event-based neuromorphic AI, announced the AKD1500 PCIe development card, an ultra-low power PCIe […]

Build your own harness, folks.

This is absolute banger paper from NVIDIA on self-evolving agent harnesses.

(bookmark it)

They introduce SoL-Pi which cuts token traffic by nearly half.

And it matches its baseline harness on GPT-5.6 Sol and Opus 5.

More details below:

Instead of tuning a harnes…

Huawei Speeds Up AI Chip Roadmap, Unveils Optical ‘Super Node’ - The Chinese tech giant moved up the release of its Ascend 960 series to 2027 as it targets data-transfer bottlenecks in massive computing clusters