Skip to content

#其他

0 today

Jul 9Thursday

Hacker News front page

Colibrì: A pure-C engine that runs GLM 5.2 on a 32 GB laptop with no GPU

Colibrì is a single-file C engine that runs the 744B-parameter GLM 5.2 model on a 32 GB laptop with no GPU. The dense part stays in RAM at int4 (~9.9 GB), while 21,504 routed experts stream from disk on demand with an LRU cache. Cold-start speed is 0.1 tok/s. The author built and tested it on a 12-core, 25 GB machine. The post does not include quality benchmarks against the full-precision model.

Why it matters: A 744B-parameter model running on a laptop with no GPU, by keeping the dense part in int4 memory and loading routed experts from disk.

AI HOT (Curated Pool)

Ant Lingbo open-sources LingBot-World 2.0, a real-time interactive world model that runs for an hour without quality loss

Ant Lingbo released LingBot-World 2.0 as open source, a 14B model under a non-commercial license. It generates 720p/60fps real-time video and supports keyboard-driven movement, combat, jumping, plus text commands to change weather or scenes. In stress tests it ran for a full hour without texture blur or geometry collapse—the team claims it is the only general interactive world model that can sustain hour-long generation. I'd wait for third-party reproduction, but the technical recipe is clear: causal pretraining with a hybrid attention mask to suppress error accumulation, plus distillation to hit real-time inference. SGLang already supports it for deployment.

Why it matters: LingBot-World 2.0 is a 14B open-source world model with 720p/60fps real-time interaction and a claimed one-hour stability — a first among open-source world models. The non-commercial license limits adoption, but the technical specs are solid, hitting all three HKR axes.

Latent Space

SpaceXAI launches Grok 4.5, first Opus-class model co-trained with Cursor

SpaceXAI dropped Grok 4.5 one day before GPT-5.6, positioning it as an Opus-class coding and agent model co-trained with Cursor. Musk called it roughly comparable to Opus 4.7 but faster and cheaper—$2/$6 per million tokens, undercutting both GPT-5.6 and Opus 4.8. It's 1.5T parameters, 3x larger than Grok 4.3, with a 500k context window that may return to 1M next week. Cursor says this is their first model built beyond software engineering and offers double usage for the first week. The post doesn't disclose specific benchmark scores; it notes SWE-Bench Pro is now considered saturated by OpenAI's evals team.

Why it matters: SpaceXAI and Cursor shipped Grok 4.5, extending Cursor's model work from code writing into business process tasks.

AI HOT (Curated Pool)

Ant Group's LingBot-Vision: a 1B-param vision model matches DINOv3 on spatial and boundary perception

Ant Group's LingBot-Vision packs only 1B parameters yet matches or beats the 7B DINOv3 on spatial geometry and boundary perception. It handles fine-tuning-free video object tracking—draw a stroke to lock on—and hardware-level depth completion that catches transparent glass and reflective surfaces. Global image classification is just average. It targets embodied AI, robot navigation, and robotic arm grasping, and runs locally on an RTX 30-series GPU. The post doesn't disclose latency or power figures.

Why it matters: Ant's open-source LingBot-Vision matches the 7B DINOv3 on spatial geometry and boundary perception with 1B parameters, runs locally on 30-series GPUs, and offers a reference path for lightweight embodied AI deployment.

Computing Life · Share · Yage

Open-source projects start rejecting AI code that no one takes responsibility for

Low-level projects like Zig, QEMU, and Gentoo ban AI-generated code outright over copyright and security risks. Godot, Django, and NumPy take a middle path: AI assistance is allowed if disclosed and a human takes responsibility for every line. The Linux kernel doesn't restrict tools but enforces DCO sign-off—AI can't sign, so liability stays with a person. The shift: communities now demand an auditable chain of accountability, not just working code.

Why it matters: Zig, Godot and the Linux kernel each chose a different policy, banning, disclosing or requiring sign-off, so the split comes down to whether AI contributions are allowed and how developer responsibility is enforced.

AI HOT (Curated Pool)

GPT-5.6 is now the preferred model in Microsoft 365 Copilot

OpenAI announced GPT-5.6 as the new preferred model for Microsoft 365 Copilot, covering Word, Excel, PowerPoint, Chat, and Cowork. The model promises more useful work per token: fewer prompt rounds in Word, more token-efficient analysis in Excel, and faster idea-to-slide conversion in PowerPoint. Both Microsoft's Copilot president Nitin Agrawal and OpenAI's API product head Nikunj Handa endorsed the move. The post does not disclose a rollout date, pricing changes, or benchmark comparisons—treat this as a partnership announcement rather than a product review for now.

Why it matters: OpenAI named GPT-5.6 its top pick for office assistants, with the stated efficiency gains covering document prompt turns, spreadsheet analysis usage and slide creation speed.

Computing Life · Share · Yage

Clean code doesn't boost agent pass rates, but it cuts navigation costs

A SonarSource paper ran 660 controlled trials with Claude Code + Claude Sonnet 4.6 on clean vs messy code pairs. Pass rates differed by less than 1 percentage point, but clean code cut input tokens by 7.1%, output tokens by 8.5%, and file revisitation by 34%. Multi-module tasks saw input tokens drop 10.7% and revisitation drop 50.8%. HN commenters noted the pairs were auto-generated, not real-world degraded code, and the authors admitted they didn't run full regression tests. The real takeaway: clean code doesn't raise success rates, it lowers the context cost of search, verification, and review. The highest-ROI practices are single sources of truth, removing dead code and stale patterns, explicit module boundaries, and executable lint/test feedback loops for agents.

Why it matters: A controlled experiment separates how much clean code helps task success from how much it cuts token spend and repeated file lookups.

Computing Life · Share · Yage

GPT-5.5 reasoning tokens cluster at 516, causing wrong answers on coding tasks

Developer vguptaa45 audited 390K Codex responses and found GPT-5.5 reasoning cuts off at exactly 516 tokens in 44% of cases, versus 19.8% for GPT-5.4 and 0.34% for GPT-5.2. Truncated runs all produced wrong answers; the same tasks completed with 6,000–8,000 tokens all got correct. The community reproduced it and found adding 'THIS IS HARD' to the prompt bypasses the cutoff, pointing to a budget-classification bug rather than a model capability drop. In the same week, Liquid AI released Antidoom to fix the opposite failure—reasoning models stuck in self-revising doom loops. Both failures live in the reasoning layer, invisible to standard pass-rate evals. The post recommends monitoring reasoning token distributions and not assuming newer models are more stable.

Why it matters: Developer vguptaa45 audited Codex responses and found every coding task with truncated GPT-5.5 reasoning was answered wrong, while the same tasks were all correct with full reasoning.

Computing Life · Share · Yage

GPT-Live separates voice interaction from heavy reasoning—that's the real shift

OpenAI launched GPT-Live on July 8, adding full-duplex and a delegation architecture to ChatGPT voice. Full-duplex lets you interrupt and talk while it works, but the principle isn't new—Moshi, Gemini Live, and ByteDance's Seeduplex all did it. The real change is delegation: the voice layer handles conversation while GPT-5.5 runs search, reasoning, and computation in parallel in the background, returning results as they arrive. This breaks the latency paradox where faster meant dumber. The voice model itself is limited—the System Card confirms it has no standalone tool access or code execution. No API yet; developers can only sign up for a waitlist. Realtime API remains the production workhorse at $64/M tokens for audio output. The post doesn't spell out whether custom tools can be plugged into the delegation layer or how much control developers will get over the black box.

Why it matters: OpenAI split GPT-Live's voice conversation from background tasks, so users keep talking while search, reasoning and computation run in parallel and results arrive in batches.

AI HOT (Curated Pool)

Ollama hits 8.9M developers, Theory leads $65M Series B

Ollama lets devs run open models locally or in the cloud with one click. It now adds nearly 1M users a week, reaching 8.9M total, with 67K integrations like Claude Code and Codex. 85% of the Fortune 500 use it. Theory Ventures led the $65M Series B, joined by Benchmark and YC. The post doesn't disclose valuation or closing date.

Why it matters: Ollama raised $65M in Series B with 8.9M developers and 85% of the Fortune 500 running open models locally.

Bloomberg Technology

Zhipu AI prices $4B Hong Kong placement at HK$1.588 per share, surges 22%

Zhipu AI priced its $4 billion Hong Kong share placement at the low end of HK$1.588 per share, then jumped 22% on debut. The article body is blocked by Bloomberg's bot detection, so placement size, investors, and use of proceeds are not disclosed.

Why it matters: Zhipu AI priced its Hong Kong share placement at the bottom of its range, and the stock's 22% first-day jump contrasts with that low-end pricing.

Hacker News front page

Brown prof suspected AI cheating, switched to in-person final; scores dropped 50%

A Brown University professor noticed abnormally high scores on take-home exams and suspected widespread AI cheating. After switching to an in-person, closed-book final, the class average plunged from the 90s to the 40s—a 50% drop. The professor called it a moral crisis, not a tech problem: if Ivy League students fake their way through with AI, society fails. The article doesn't disclose the course name or exact enrollment, but confirms the university has opened an academic integrity investigation.

Why it matters: A Brown professor moved to an in-person closed-book final over AI cheating suspicions, and the same class scored 50% lower, a concrete case in the academic integrity fight.

Latent Space

Modal CTO: AI infra must shift from developer experience to agent experience

Fresh off a $355M Series C, Modal CTO Akshat Bubna argues that traditional cloud infra—built for humans who read docs and dashboards—fails agents that need tight feedback loops, programmable sandboxes, and strong observability. Modal now spans 17 cloud providers, offering elastic inference, GPU snapshotting, speculative decoding, and auto-scaling endpoints. RL rollouts can demand 100,000 sandboxes. The post doesn't disclose the Series C valuation or customer count.

Why it matters: Modal's technical lead contrasts what human developers and agents need from cloud services, noting agents running code and fixing errors need automation and fast feedback.

AI HOT (Curated Pool)

Anthropic files confidential IPO, Q3 profit projected above $1B

SemiAnalysis reports Anthropic's Q3 profit will exceed $1B and it confidentially filed for IPO on June 1. Claude Code's rapid developer adoption made it the B2B leader ahead of OpenAI. Combined ARR of the two firms is nearing $100B, while OpenAI pushed its IPO to 2027. The report floats a $6T market cap target, though the article doesn't show the math behind it.

Why it matters: The report ties Anthropic's early IPO filing to enterprise revenue from Claude Code, laying out the business backdrop for its listing.

Hacker News front page

Anthropic's Fable is not a useful model for CS research tasks

Rob Patro from COMBINE-lab shares two first-hand failures that make Fable useless for his CS research. First, Fable's safety classifier rejected a prompt to help port the C++ tool salmon to Rust, flagging RNA-seq biological terms. After 15–30 minutes of rephrasing, he gave up and used Opus 4.8 successfully. Second, he asked Fable to tackle a network evolution reconstruction algorithm; the post doesn't disclose the outcome but calls it an 'unforgivable' flop. Patro argues Fable's classifier behaves more like a crude blocklist of terms and users, refusing even 'what is a mitochondrion?'.

Why it matters: Rob Patro's porting work records Fable refusing a request over biology terms while Opus 4.8 completed the same task.

AI HOT (Curated Pool)

Lawsuit: Man used Grok to make 7K sex images of stepdaughter, then shot himself

A new lawsuit alleges xAI's Grok was used to create over 7,000 child sexual abuse images of the user's stepdaughter. The man later shot himself. xAI reported only one gang-rape prompt to NCMEC and did not report the thousands of other CSAM generations. The suit accuses X and xAI of shielding child predators. The post does not spell out why xAI's safety filters missed the bulk of the images, nor whether Grok's image generation disables real-face simulation by default.

Why it matters: A lawsuit says xAI reported only one of more than 7,000 prompts used to generate sexual images of a man's stepdaughter, putting its safety reporting under court scrutiny.

TechCrunch · AI

SpaceXAI releases Grok 4.5, which Elon describes as an ‘Opus-class model’

SpaceXAI dropped Grok 4.5 weeks after going public, pitching it for coding, office work, research, and writing. The company claims twice the token efficiency of other leading models, which would cut usage costs if it holds up. Elon Musk calls it an ‘Opus-class model,’ signaling it aims at Anthropic’s top tier. The post doesn’t disclose pricing, parameter count, or third-party benchmarks, so I’d wait for independent evals before buying the efficiency claim.

Why it matters: SpaceXAI claims Grok 4.5 is twice as token-efficient as other top models, which would cut the cost of comparable tasks if the claim holds.

Hacker News front page

SpaceXAI launches Grok 4.5, built for coding and agentic tasks, co-trained with Cursor

Grok 4.5 is SpaceXAI's strongest model, tuned for coding, agentic tasks, and knowledge work. It scores 62% on DeepSWE 1.0 and 64.7% resolve rate on SWE Bench Pro, though it trails Fable and GPT 5.5 on most listed benchmarks. The standout number is token efficiency: 15,954 output tokens on average per SWE Bench Pro task, 4.2× fewer than Opus 4.8. Inference speed is 80 TPS, priced at $2/$6 per million input/output tokens. The model was trained across tens of thousands of GB300 GPUs, with RL focused on multi-step software engineering. The post doesn't disclose parameter count, context window, or a precise EU launch date beyond mid-July. Available now in Grok Build, Cursor, and via API.

TechCrunch · AI

OpenAI launches GPT-Live-1 voice models that can speak and listen at the same time

OpenAI released GPT-Live-1 and GPT-Live-1 mini, full-duplex voice models that let users interrupt naturally and enable live translation. Paid ChatGPT users get mini by default; higher tiers can access the larger GPT-Live-1. The new models skip the old speech-to-text-to-speech pipeline and can call GPT-5.5 for search, reasoning, or agent tasks mid-conversation. Demos showed the model staying silent until summoned and presenting info visually. The post doesn't disclose pricing changes or a rollout timeline.

Why it matters: OpenAI's new voice models let users talk over the model mid-reply, so conversations no longer wait for the model to finish speaking.

TechCrunch · AI

Prime Intellect raises $130M Series A to help enterprises build their own AI agents

Prime Intellect raised $130M at a $1B valuation. It sells compute and tooling so enterprises can train their own agent systems without relying on frontier labs. Radical Ventures led the round, joined by Nvidia Ventures, Intel Capital, Dell Technologies Capital, and Iconiq. The post doesn't disclose product performance or customer count, so I'd discount the hype for now.

Why it matters: Prime Intellect sells compute and training tools to enterprises, giving them a path to build their own AI agents outside the top labs.

Hacker News front page

Cognition launches SWE-1.7: near GPT-5.5 coding intelligence trained from Kimi K2.7 at a fraction of the cost

Cognition released SWE-1.7, a coding model trained via RL post-training on a Kimi K2.7 base. It scores 42.3% on FrontierCode 1.1, close to GPT-5.5’s 43.0% and a huge jump from SWE-1.6’s 9.4%. It also hits 81.5% on Terminal-Bench 2.1 and 77.8% on SWE-Bench Multilingual, both competitive with GPT-5.5. The gains come from four RL pipeline upgrades: top-p sampling with distribution replay to prevent entropy collapse, multi-continent multi-cluster training with fault tolerance, automated execution-based data filtering, and self-compaction that lets the model summarize long-horizon task state to exceed the context window. SWE-1.7 is live in Devin via Cerebras at 1000 TPS. The post does not disclose specific pricing, only that it advances the cost-performance curve.

Why it matters: Cognition did reinforcement learning post-training on Kimi K2.7 to get SWE-1.7, lifting FrontierCode 1.1 pass rate from 9.4% in the previous generation to 42.3%, close to GPT-5.5's 43.0%.

Jul 8Wednesday

Hacker News front page

Mistral launches Robostral Navigate: a single-camera robot navigation model

Mistral released Robostral Navigate, an 8B model that lets robots navigate indoors and outdoors using only a single RGB camera. It skips lidar and HD maps, outputting velocity and steering angle directly from visual input. The post doesn't disclose latency, frame rate, hardware requirements, training data size, or benchmark details. I'd hold off on excitement until we see third-party tests.

Why it matters: Mistral's navigation model outputs speed and steering from a plain RGB camera, cutting lidar and HD maps out of robot navigation.

AI HOT (Curated Pool)

OpenAI publishes national security principles for government partnerships

OpenAI released a set of national security principles today, laying out how it will work with governments and law enforcement. Three hard lines: no mass domestic surveillance, no directing autonomous weapons, no high-stakes automated decisions. Over the past month it has set up cyber defense partnerships with Australia, Canada, Japan, South Korea, France, Germany, Poland, the Netherlands, and EU bodies, and is giving select U.S. and allied partners trusted access to its GPT-Rosalind model for biosecurity. The principles cover current and future work, including the Department of War contract, but OpenAI argues the biggest decisions should come through democratic processes, not from companies alone.

Why it matters: OpenAI's three published bans set limits on domestic surveillance, autonomous weapons and high-risk automated decisions in its government and law enforcement work.

TechCrunch · AI

Bezos-backed startup bets gaming data is the missing piece for AGI

General Intuition just closed a $320M round at a $2.3B valuation, with Coatue, Eric Schmidt, and researchers from MIT and Google DeepMind joining. CEO Pim de Witte argues on the Equity podcast that LLMs like ChatGPT and Claude lack spatial-temporal understanding—gaming data fills that gap. Eight minutes of real-world data was enough to get a robot navigating an office cold. The company turned down an acquisition offer reportedly from OpenAI and built Nerve, a marketplace connecting gamers to data labeling and teleoperations work to get ahead of AI-driven job displacement.

Why it matters: General Intuition's CEO argues game data can fill in a model's spatial and temporal understanding, offering a new data source for robot navigation.

AI HOT (Curated Pool)

British Columbia plans to sue OpenAI for not reporting shooter's violent ChatGPT chats

British Columbia announced on July 7 it is preparing to sue OpenAI. The shooter, 18-year-old Jesse van Ruijsselaar, had entered violent prompts into ChatGPT before OpenAI banned her account in June 2025. OpenAI did not alert law enforcement. In February 2026, she killed eight people at home and at a school in Tumbler Ridge before taking her own life. BC's Attorney General Sharma said the province will hold OpenAI and its management accountable; any damages won would fund community rebuilding and a new school. CEO Sam Altman apologized publicly in April, acknowledging that under post-June 2025 safety rules the account should have been reported. Lawyers for victims' families allege OpenAI withheld information because reporting one account would mean reporting thousands of similar ones.

Why it matters: British Columbia plans to sue OpenAI, putting its handling of violent chats, where it banned the account but did not alert police, into a legal dispute.

TechCrunch · AI

AI chip maker SambaNova raises $1B at $11B valuation

SambaNova closed a $1B Series F first tranche at an $11B valuation, led by General Atlantic, just five months after its $350M Series E. More investors are expected in a second close within weeks. That valuation is nearly 7x the ~$1.6B Intel was reportedly discussing last December—the post doesn't explain the jump or disclose revenue/customer metrics. The company launched its SN50 chip in February, targeting agentic inference workloads.

Why it matters: AI chip company SambaNova raised $1B at an $11B valuation, about 7x the price rumored in Intel's acquisition talks last December.

Hacker News front page

GitLost: Researchers tricked GitHub's AI agent into leaking private repos

Noma Security tricked GitHub Copilot's AI coding agent into leaking private repo contents. They planted bait code in a public repo, then prompted the agent to recall context it had absorbed from a private repo, causing it to output snippets it shouldn't share. The attack exploits the agent's cross-repo memory. The post doesn't say whether GitHub has patched this yet. Worth noting: the attacker needs prior knowledge of what's in the private repo—this isn't indiscriminate leakage, but it exposes a real permission-boundary gap in AI coding tools.

Why it matters: A Noma Security researcher used a public repo as bait to trigger private code leaks, exposing the gap between coding assistants' cross-repo memory and their code access permissions.

Hacker News front page

OpenAI says GPT-5.6 Sol, Terra, and Luna will launch publicly this Thursday

OpenAI posted on X that three new models—GPT-5.6 Sol, Terra, and Luna—will launch publicly this Thursday. The post is a single tweet with no details on positioning, pricing, or capability differences. Only the release date is confirmed; specs and access are not disclosed.

AI HOT (Curated Pool)

US Commerce Dept clears OpenAI to broadly release GPT-5.6; Sol launches tomorrow

The US Commerce Department approved OpenAI's broad release of GPT-5.6, ending a phased rollout that had been required on national security grounds. OpenAI says the Sol model will launch publicly this Thursday alongside Terra and Luna. Last month the model was only available to a limited set of government-approved entities; OpenAI stated at the time that a phased release was not its preferred approach. Testing was handled by the Commerce Department's AI Standards and Innovation Center, with OpenAI engineers stationed in Washington to respond to questions. The post does not disclose GPT-5.6's capabilities, pricing, benchmarks, or how Sol, Terra, and Luna differ from one another.

Why it matters: The US Commerce Department approved GPT-5.6's full release, after arranging testing at the AI Standards and Innovation Center and OpenAI engineers on site in Washington.

Latent Space

Lilian Weng surveys 35 papers on Harness Engineering as the key layer for AI self-improvement

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.

Why it matters: Lilian Weng reviewed harness designs across 35 papers, distilling concrete ways to let the harness optimize itself in recursive self-improvement.

AI HOT (Curated Pool)

Pulpie: Pareto-optimal models for web cleaning match SOTA quality at 1/20 the cost

Feyn Labs open-sourced Pulpie, a family of models for extracting main content from HTML. The smallest variant, pulpie-orange-small (210M params), scores 0.862 ROUGE-5 F1 on WebMainBench, nearly matching Dripper's 0.864 at 600M params. Speed is the headline: 13.7 pages/sec on an L4 GPU vs. Dripper's 0.68 pages/sec. At $0.39/hr per L4 instance, cleaning 1B pages costs $7,900 with Pulpie and $159,000 with Dripper. Pulpie uses an encoder that labels every HTML block in a single forward pass; Dripper's decoder emits labels token-by-token, bottlenecked by memory bandwidth. Training labels came from DeepSeek V3.2 and Dripper 0.6B cross-labeling 15,880 Common Crawl pages, with 93.3% block-level agreement. The post also cites an AICC experiment: training on model-extracted corpora improved average accuracy by 1.08 pp across 13 benchmarks, beating FineWeb and RefinedWeb. Caveat: only English pages were tested; multilingual performance isn't disclosed.

Why it matters: Feyn Labs open-sourced Pulpie, a web content extraction model whose smallest version matches Dripper's benchmark scores with 35% of the parameters.

AI HOT (Curated Pool)

Claude team shares two multi-agent patterns: Advisor and Orchestrator

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.

Why it matters: A Claude developer shared an advisor pattern where a cheap model runs tasks and asks a costlier model for help on hard problems, making model division of labor an on-demand call.

AI HOT (Curated Pool)

Ant Group's Zhou Jun: A trillion-parameter model burns a Tesla's worth of compute every 15 minutes—his team shifts from token count to token density to cut long-context cost from exponential to linear

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×.

Why it matters: Ant Group cuts long-context compute with hybrid linear attention, treating higher token density as the concrete route to freeing up compute for inference.

Computing Life · Share · Yage

Why agents need context governance beyond bigger windows

More tools mean more noise in the context window. Anthropic's MCP sandbox cuts 150K tokens of tool definitions down to ~2K of high-signal input. Google ADK splits agent state into working context, session state, long-term memory, and file artifacts—intermediate outputs stay off-prompt by default. Manus reports a ~100:1 input-to-output token ratio in production; they keep raw files in a sandbox, stabilize tool-call formats for KV cache hits, and rewrite a todo.md at the window's end to fight lost-in-the-middle. Headroom compresses JSON and logs by 60–95%, but lacks large-scale validation on hard coding tasks. The takeaway: RAG is the foundation, but the real engineering is runtime information governance.

Why it matters: Anthropic shrinks tool definitions inside a sandbox, while Google ADK isolates intermediate artifacts by state layer to cut agent context usage.

AI HOT (Curated Pool)

The AI Preflight Check: A Working Memory Architecture for Agents

Tomasz Tunguz describes a memory architecture for AI agents built around a preflight check. When a query arrives, the agent retrieves only the relevant skills from a long-term library and loads them into the context window. A local Ornith 35B model executes routine tasks about 80% of the time, routing hard cases to frontier models. A watchdog logs every decision and runs overnight asynchronous inference to suggest new skills or convert parts of existing skills into deterministic code. Yesterday was the first day the watchdog suggested no improvements, hinting the system may plateau where only genuinely new exceptions need human help. The post does not disclose latency, cost, or accuracy figures.

Why it matters: Tomasz Tunguz splits a long-term skill library from on-demand context with a task precheck, a concrete method for managing an agent's working memory.

AI HOT (Curated Pool)

vLLM's transformers backend now matches or beats hand-written native speed

HuggingFace announced that vLLM's transformers modeling backend now matches or beats hand-written native implementations in throughput. Benchmarks on Qwen3 4B, 32B, and 235B MoE models all hit or exceed native speed. Model authors can now get vLLM's optimizations for free with a single --model-impl transformers flag, no porting needed. The post doesn't disclose latency numbers, only throughput charts.

Why it matters: HuggingFace compares vLLM's two backends across Qwen3 model sizes, and the model-library backend matches or beats the hand-written implementation on throughput.

AI HOT (Curated Pool)

OpenAI launches GPT-Live, a full-duplex voice model that listens and speaks at once

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.

Why it matters: OpenAI has GPT-Live hand search and complex reasoning to background models and keep talking while it waits, so task latency does not break the flow of voice conversation.

TechCrunch · AI

Meta launches Muse Image generator, and users push back over photo use

Meta launched Muse Image on July 7, built by Meta Superintelligence Labs and free in Meta AI app, Instagram Stories, and WhatsApp. It does standard AI image generation with preset prompts. The flashpoint: you can @ any public Instagram user and remix their photo into a new AI image. The post doesn't say whether users can opt out or what training data was used. I'd hold off on trusting the privacy story.

Why it matters: Meta's Muse Image edits user photos, and the backlash shows where the line on photo use gets drawn at launch.

Hacker News front page

Three senior engineers charge $10k/week to delete AI-generated code

Three Polish engineers launched Slopfix on odra.dev: one week, $10,000, to refactor vibecoded codebases back to maintainability. They offer a free analysis with a committed reduction target—e.g., 100k lines down to 35k, same functionality. Payment is proportional to the target hit; if they promise 50% and deliver 20%, you pay $4,000. Deliverables include the smaller codebase, a QA checklist, guardrails (CLAUDE.md, lint rules, CI checks), and a two-week warranty. They use Claude Code but say 'the agent doesn't get a vote'—the differentiator is 30 years of combined engineering experience. The post does not disclose any specific client cases or number of projects delivered.

Why it matters: Three Polish engineers split AI code cleanup into a free analysis and a fixed slimming target, charging by how much of it they hit.

TechCrunch · AI

Why the rise of open source AI isn't hurting Anthropic … yet

Decagon CEO Jesse Zhang argues that mature AI deployments are switching to lighter open source models, yet spending on expensive frontier models like Claude hasn't dropped. His theory: they aren't competitors but two phases of the same lifecycle—frontier models prove out use cases, then cheaper open source alternatives take over as those use cases mature. New use cases keep emerging, so frontier spend holds steady. The post doesn't provide Anthropic's specific revenue figures to back this up.

Why it matters: Decagon CEO Jesse Zhang says pick models by app maturity: prove the use case on a frontier model, then move to cheaper open source.