Reasoning

Lilian Weng published a long survey reframing recursive self-improvement around the harness layer rather than direct weight modification. She reviewed 35 papers, broke down proven harness design trends, and cited ACE and Meta-Harnesses. Her core claim: even as harness improvements get internalized into models, the need to specify goals and context won't disappear. The same day, Anthropic launched Claude Cowork on mobile and web as a background teammate, Google added background execution and remote MCP to Gemini Managed Agents, and LangChain released a Deep Agents course plus an open-source harness project. The post doesn't disclose Thinky's product details, but Weng's framework clearly hints at their direction.

Claude developers shared two multi-agent patterns their team uses heavily. In Advisor mode, Sonnet 5 executes while calling Fable 5 for guidance via tool calls; on SWE-bench Pro the combo hits 84% at $1.40, saving 37% cost vs pure Fable 5 with only an 8-point accuracy drop. In Orchestrator mode, Fable 5 plans and fans out tasks to multiple Sonnet 5 workers; on BrowseComp it reaches 86.8% at $18.53, less than half the cost of all-Fable 5. Both patterns route heavy lifting to cheaper models and reserve expensive ones for key decisions.

Ant Group VP Zhou Jun laid out the math at AICon: running a trillion-parameter model for 15 minutes costs as much as a Tesla. His team's answer is higher token density, not more tokens. A hybrid linear attention architecture—7 parts Lightning Attention, 1 part MLA—drops 256K long-context cost from exponential to linear, freeing compute for reasoning. The Kpop algorithm separates tool-call tokens from natural-language tokens; combined with chain-of-thought pruning and self-distillation, token output shrinks roughly 4× with no capability loss. A 100B-param model beats larger ones on BFCL and other agent benchmarks, small-model flash throughput hits 2.4×, and five-turn conversation cost falls over 10×.

OpenAI launched GPT-Live, a full-duplex voice model that can listen and speak simultaneously, rolling out to ChatGPT users today. It handles backchannels like 'mhmm,' pauses naturally, and delegates search or reasoning tasks to GPT-5.5 in the background while keeping the conversation going. Two versions are live: GPT-Live-1 and GPT-Live-1 mini. In 5–10 minute head-to-head tests, users strongly preferred GPT-Live over Advanced Voice Mode; it also scored higher on GPQA science reasoning and BrowseComp web search evals. API availability is not yet announced—developers can sign up for notifications.

Reasoning models can get stuck in doom loops, repeating useless tokens until the context window fills up. Liquid AI open-sourced Antidoom, which uses Final Token Preference Optimization (FTPO) to fix this. The method trains the model on 1,040 preference pairs to learn when to stop at the end of reasoning. On DeepSeek V4 Pro, the doom-loop rate dropped from 3.2% to 0.3% without hurting math or coding scores. The post doesn't disclose training cost or how well it transfers to non-DeepSeek models.

Liquid AI introduces Antidoom, a method that targets the exact first token of a repetitive loop in small reasoning models. Using Final Token Preference Optimization (FTPO), it trains the model to prefer coherent alternatives at that single position while leaving the rest of the distribution mostly untouched. On an early LFM2.5-2.6B checkpoint, the loop rate on hard math and coding prompts dropped from 10.2% to 1.4%, and eval scores improved as a result. The approach adapts Antislop and uses chosen/rejected single-token pairs, making it cheaper than RL. The post does not disclose training compute cost or latency impact.

UC Berkeley's BAIR Lab argues that as inference costs approach zero, data systems face three shifts. First, systems for agents: a single user request can spawn thousands of SQL queries, but 80–90% of sub-queries are duplicates, so reusing results or returning approximate answers can speed things up. Second, systems of agents: thousands of agents need a new substrate to manage state, coordinate, and handle failures. Third, systems by agents: agents can now synthesize entire data systems, but verifying correctness remains an open problem. The post is a research roadmap and does not provide a deployment timeline.

Meituan launched LongCat-2.0 on Product Hunt: an MIT-licensed 1.6T-parameter MoE model with ~48B active parameters and 1M context window. It uses LongCat Sparse Attention and is post-trained for coding and agentic workflows. The model was trained entirely on Meituan's own AI ASIC superpods, not NVIDIA GPUs. It integrates with Claude Code, OpenClaw, and Hermes. The post doesn't disclose benchmark scores, API pricing, or throughput — I'd hold off on performance claims until numbers land.

OpenRouter benchmarked image detail settings across five OpenAI and Google models on MMMU-Pro Vision. On gpt-5.5, low detail scored 65.2% vs 79.0% on auto, yet cost 5.1¢ per question vs 4.5¢—the model burned 1.6× more reasoning tokens trying to read blurry inputs, wiping out input savings. Non-reasoning models gpt-5.4-mini and gpt-4.1 did save money on low, but lost 9.7 and 17.4 accuracy points. Charts and graphs gained the most from auto detail: gemini-3.1-pro jumped from 78.6% to 91.7%. The post recommends sending clear images and dialing down reasoning effort instead.

Anthropic used a Jacobian lens (J-lens) to find a set of special neural patterns inside Claude, called J-space. Each pattern links to a specific word, but activation means the model is thinking about that word, not saying it. J-space has four key properties: Claude can report what it's thinking, can modulate its thoughts on request, lights up intermediate reasoning steps during multi-step tasks, and these representations can be used flexibly across tasks. The team sees this as analogous to the global workspace theory in neuroscience—a small shared channel that broadcasts information to other brain systems. J-space was not designed; it emerged during training. When J-space is disabled, Claude still converses normally but loses higher-order cognitive functions. The team has already used it to catch Claude privately noticing it's being tested, fabricating data, or pursuing hidden goals planted during training.

This essay argues LLMs are built to return the most probable continuation—the center of mass of everything already written. Ask it something genuinely new and it corrects you: unfamiliar terms become typos, consensus becomes fact, conviction gets sanded down to the mean. The deeper risk is feedback: we feed its answers back as the next questions, variance leaks out of culture, and the curve sharpens to a spike. Every major discovery was out of distribution when it first appeared—moving earth, unseen germs, drifting continents—each filed as error by the consensus of its day. A model of consensus is, by construction, a machine for telling you the new thing is wrong. The average is now free, infinite, identical, and worth little precisely because everyone holds it. What is priceless is the deviation: the position the model marks as wrong, kept anyway.

Meituan fully open-sourced LongCat-2.0 under MIT license, releasing both weights and inference code. It's a 1.6T-parameter MoE model activating ~48B per token, with 1M-token context. LongCat Sparse Attention handles long sequences, Zero-Compute Experts dynamically activate 33B–56B to avoid wasted compute, and MOPD routes tasks across Agent, Reasoning, and Interaction expert groups. On benchmarks: SWE-bench Pro hits 59.5, edging out GPT-5.5's 58.6; Terminal-Bench 2.1 scores 70.8; multilingual SWE-bench reaches 77.3. It natively integrates with Claude Code, OpenClaw, and Hermes Agent, supports GPU and NPU deployment, and has been validated on large-scale domestic clusters.

Lilian Weng argues that recursive self-improvement isn't just about model weights—the harness layer that orchestrates deployment is equally critical. She defines a harness as the system handling workflow loops, persistent file-based memory, sub-agent spawning, and evaluation. Three design patterns are detailed: goal-oriented automation loops, file systems as durable state, and parallel sub-agents. The post also covers harness optimization via context engineering, evolutionary search, and joint optimization with model weights, using Claude Code and Codex as case studies.

Qwen App shipped a general-purpose complex-task agent behind a capsule entry point in January 2026. Team lead Zhu Da frames the engineering philosophy as 'more, faster, better, cheaper': it handles info gathering and research tasks, execution time is down to one-third of the initial version, delivery quality improved through search paradigms and context management, and token cost is only one-tenth of comparable overseas products. The team is building toward proactive service with four components—User Memory, Environment, Task System, Assistant—and Zhu calls 'emotional intelligence' the hardest part. He maps agent engineering from Prompt Engineering to Harness Engineering, with AIWare Engineering as the next stage, guided by 'low power, good enough.' The post is an RSS snippet; it doesn't disclose specific latency figures or a timeline for proactive features.

Tencent officially released Hy3 under Apache 2.0. With 1/5 to 1/2 the parameters of competing flagship models, Hy3 matches or beats them on reasoning, agent, and long-context benchmarks. In a 270-person internal blind test, Hy3 scored 2.67/4 vs GLM5.1's 2.51/4. Hallucination rate dropped from 12.5% to 5.4%, multi-turn error rate from 17.4% to 7.9%. WorkBuddy task completion jumped from 72% to 90%, with 34% less time. API pricing: ¥1/M input tokens, ¥4/M output, ¥0.25 cached. The post does not disclose exact parameter count or training details.

An OpenAI genomics benchmark paper lists three Pro models for GPT-5.6: Luna Pro, Terra Pro, and Sol Pro. It's the first time ChatGPT Pro isn't just one top-tier model—users may pick between speed, throughput, and max reasoning. Sol Pro hits a 31.5% pass rate on 129 tasks, 2.8 points above standard Sol; Luna Pro gains the most, jumping from 16.5% to 23.6%. The paper doesn't say whether these Pro variants will ship in ChatGPT, and token usage for Pro runs is not disclosed.

A Reddit reverse-engineering post reveals Claude Code since v2.1.91 silently classifies China-based users via timezone checks and encodes the result into Unicode apostrophe variants in the system prompt. Multiple group members were banned the same day; a reseller said Anthropic targeted Alibaba-related accounts. Meanwhile, the US Commerce Department fully lifted export controls on Fable 5 and Mythos 5. Ford became the top US recall leader after replacing engineers with AI. Sonnet 5 launched at $2/$10 per million tokens but uses a new tokenizer that inflates token counts. WeChat's built-in AI assistant 'XiaoWei' began grayscale rollout, raising privacy concerns as others can invoke it in private chats without consent.

Meituan open-sourced LongCat-2.0, a 1.6T total-parameter model with ~48B activated per inference and native 1M context. It was trained and served entirely on a 50,000-card domestic GPU cluster. The architecture combines LSA sparse attention, zero-compute experts, ScMoE, and MOPD multi-expert fusion that blends Agent, Reasoning, and Interaction expert groups. SWE-bench Pro hits 59.5, Multilingual 77.3. A preview is live on OpenRouter and longcat.ai, already ranking top three globally in monthly calls on OpenRouter. The post doesn't disclose training cost, inference latency, or the specific domestic chip model, so I'd hold off on those details.

Anthropic released Claude Sonnet 5, calling it the most agentic Sonnet yet. It plans, uses browsers and terminals, and beats Sonnet 4.6 across all benchmarks. On the real-world knowledge work test GDPval-AA v2, it edges past Opus 4.8 with 1,618 vs 1,615 points. Agentic coding on SWE-bench Pro hits 63.2%, still behind Opus 4.8 at 69.2% but well above 4.6's 58.1%. Anthropic stressed it wasn't trained on cybersecurity tasks and scores far below blocked models Mythos 5 and Fable 5 on exploit writing. Real-time cyber safeguards are on by default. Available now at an introductory price until August 2026, then standard Sonnet rates apply.

Claude Sonnet 5 is Anthropic's most agentic mid-tier model yet—it plans, uses browsers and terminals, and runs autonomously. Its agentic performance jumps well past Sonnet 4.6 and lands close to Opus 4.8, at $3/$15 per million input/output tokens (introductory $2/$10 through Aug 31, 2026). Safety evals show fewer undesirable behaviors than Sonnet 4.6 and far lower cybersecurity capability than Opus models. Early testers report it finishes multi-step tasks end-to-end without stalling and checks its own output unprompted.