Skip to content

#其他

0 today

Jul 3Friday

AI HOT (Curated Pool)

Zuckerberg tells staff AI agents aren't progressing as fast as he'd hoped

At an internal town hall Thursday, Meta CEO Mark Zuckerberg said AI agent development hasn't accelerated the way executives expected. Earlier this year Meta laid off ~8,000 employees and reassigned ~7,000 to AI groups. Zuckerberg admitted the cuts weren't as 'clean' as they should have been and the upside of the new AI-focused structure hasn't materialized yet. He expects improvements from AI investments in the next 3–6 months. Meta is on track to spend up to $145 billion on AI infrastructure this year.

Why it matters: Zuckerberg told a Meta all-hands that AI agent development is slower than executives expected, and named a 3-to-6-month window for the investment to show improvement.

Latent Space

Adobe experiments with agentic sites that assemble pages per visitor in real time

Adobe Principal Scientist Carlos Sanchez demoed 'agentic sites' at AIEWF: the system infers visitor intent from browsing and search signals, retrieves from existing company content, and assembles a page in real time. A camper searching for coffee saw a product page reorganized around outdoor brewing. Sanchez says this works today, with 1–2 second latency and ~1–2 cents per page in inference cost. Adobe hasn't deployed it on production customer sites yet and is looking for early experimenters. The post doesn't name the underlying model or give a rollout timeline.

Why it matters: Adobe's chief scientist demoed an agentic site that assembles pages in real time from visitor intent, rebuilding one product page around an outdoor coffee theme, a transferable approach to personalized content.

AI HOT (Curated Pool)

Microsoft launches $2.5B Frontier Company to embed 6,000 AI engineers at enterprise clients

Microsoft created a new unit called Frontier Company with a $2.5B budget and 6,000 industry and engineering experts embedded at customer sites, targeting measurable business outcomes. Rodrigo Kede Lima leads it; Microsoft Commercial CEO Judson Althoff calls it the industry's largest results-oriented engineering org. Microsoft positions itself as platform-neutral versus OpenAI and Anthropic, which deploy only their own models—ironic coming from Microsoft. System integrators Accenture, Capgemini, EY, KPMG, and PwC will help scale globally. OpenAI's DeployCo raised over $4B and fields roughly 150 on-site engineers; Anthropic partnered with Blackstone and Goldman Sachs for a deployment firm aimed at mid-sized companies. All three now agree: real AI value requires weaving into existing business processes, data pipelines, and compliance—not just shipping a chat tool.

Why it matters: Microsoft set up Frontier Company, putting $2.5 billion and 6,000 experts inside customer companies and billing on business outcomes, a look at AI delivery shifting from selling tools to on-site work.

Hacker News front page

The Short Leash Method: Using AI agents for security-critical code without sacrificing quality

Greg Slepak from okTurtles outlines a method where the developer stays in the loop at all times—reviewing every diff, denying permissions when the agent veers off, and committing after each subtask. He rejects fully autonomous 'vibe engineering' and argues that even non-frontier models can beat Fable 5 this way. For reviews, AI acts as a fast linter while the human catches directional issues; the PR author must do a line-by-line self-review and disclose which model was used.

Why it matters: Developer Greg Slepak shares a short-leash method: commit after each subtask and approve a diff before the model touches a file, the opposite of parallel multi-model vibe coding.

AI HOT (Curated Pool)

LMSYS shares how AI agents are used to speed up SGLang development

The SGLang team turned recurring dev workflows—benchmarking, profiling, CUDA crash debugging, adding diffusion pipelines—into executable SKILL files that agents follow. The repo now includes skills for debugging, integration, and CI, with a separate skill set for diffusion models. For performance, profiler skills produce fixed kernel tables, overlap-opportunity tables, and fuse-pattern tables; KDA-Pilot automates B200 kernel task comparison and correctness checks, with three PRs already merged. They also built a SOTA performance loop that breaks chasing the latest numbers into fair benchmarking, gap analysis, profiling, patching, and revalidation, adding external review via Humanize/RLCR and lower coordination cost via Codex Goal. The post warns that agents generate more plausible-looking changes that still need careful review—developers should focus on defining problems, picking evidence, and deciding what ships.

Why it matters: The SGLang team turned repeated work like benchmarking, profiling and CUDA crash debugging into SKILL files agents can run, and published how the debugging, integration and CI skills are organized.

AI HOT (Curated Pool)

Anthropic and Pentagon clash over Claude military guardrails; DoD already switched two-thirds of usage

WSJ court filings reveal months of emails between Anthropic CEO Dario Amodei and Pentagon deputy Emil Michael. The core dispute: Anthropic wants to ban fully autonomous weapons and certain surveillance uses for Claude; the Pentagon wants the model available for all lawful national security scenarios. Michael said he wouldn't 'force it' if the gap was too wide. The Pentagon then labeled Anthropic a supply-chain risk and blocked partners from using its models on DoD projects. A judge paused some restrictions; the government is appealing. Michael stated two-thirds of operations previously using Anthropic have already switched to other AI tools.

Why it matters: Court filings disclose email disagreements between Anthropic and the Pentagon over guardrails for military use of Claude, with both sides' red lines laid out in the correspondence.

Financial Times · Technology

Altman’s AI safety proposal: let us win, or everybody loses

Sam Altman published an FT op-ed arguing that AI safety depends on letting a few trusted labs like OpenAI win. He claims open-source and decentralized development risk putting dangerous capabilities in the wrong hands, so regulation should concentrate resources on a small number of vetted entities. The piece offers no concrete safety standards or external oversight mechanisms. It reads more like a public pitch for centralization dressed in safety language.

Why it matters: Altman argues in the FT that AI safety should sit with a few trusted labs, and his limits on open and decentralized work show how top companies make the public case for concentrating regulatory power.

Jul 2Thursday

AI HOT (Curated Pool)

Kuaishou's Kling AI secures $2.028B from initial investors at $18B post-money valuation

Kuaishou confirmed that Beijing Kling signed agreements with 21 initial investors for a $2.028B cash injection. On the same day, 15 additional investors joined with another $766M. The round is capped at $3B for roughly 16.67% equity, giving a $18B post-money valuation—down from the $20B target Kuaishou set in April. A source close to the deal said Kuaishou plans to list Kling AI in Hong Kong within 12 months, with IPO proceeds earmarked for compute, data centers, and talent.

Why it matters: Kuaishou's Kling closed a first round of $2.028 billion at an $18 billion post-money valuation, $2 billion under its own April target, giving video generation's leading project a concrete funding benchmark.

TechCrunch · AI

Sam Altman proposed OpenAI donate 5% equity to a US sovereign wealth fund

Sam Altman proposed donating 5% of OpenAI's equity to a US sovereign wealth fund, with other AI companies contributing similar stakes. The goal is to ease political friction and public backlash over AI profits. Trump confirmed related talks in June but gave no numbers. The post doesn't spell out who would run the fund, how equity converts to cash, or whether other firms agreed.

Why it matters: Sam Altman proposed that OpenAI donate 5% of its equity to a not-yet-created US sovereign wealth fund and urged other AI companies to follow; Trump confirmed talks in June without giving a number.

Latent Space

Paul Bakaus on skill engineering and why one-shot AI design is a dead end

Paul Bakaus presented Impeccable at the AI Engineer World’s Fair, an open-source design skill system for coding agents. Instead of one-shot full-site redesigns, users steer output with terms like 'bolder' or 'quieter' that the skill translates into precise design actions. Bakaus calls this 'skill engineering'—compressing expert vocabulary so agents don't converge on generic results. He noted designers now make up at least half of Impeccable's audience, using it as a bridge into code. He rejects full auto mode, arguing the goal is to insert human judgment at the exact point it matters most.

Why it matters: Paul Bakaus open-sourced Impeccable, a design skill pack that has coding agents edit step by step in a designer's own words instead of rebuilding a whole site at once.

AI HOT (Curated Pool)

Citi, Adobe, and others restrict flagship AI models as usage-based billing triples costs

404 Media obtained internal chats and emails from six companies including Atlassian, Adobe, and Amazon, showing a wave of restrictions on expensive AI models. Citi blocked access to Claude Opus 4.6/4.7 and GPT-5.5, directing staff to use GPT-5.3-Codex or Claude Sonnet 4.6 for routine work. Atlassian's monthly AI spend jumped from $5M to over $15M, with a full-year forecast above $120M; employees now burn through their token quota in two to three days. Adobe let its unlimited Claude agreement expire, and Amazon quietly added usage caps after removing an internal leaderboard. Several entertainment firms also blew through their token budgets—one developer consumed nearly half the company's allocation. Accenture found that heavy token usage often came from converting PDFs to slide decks, not coding, and is now selling 'token cost economics' consulting. GitHub plans to switch to open-source models and is testing per-person billing.

Why it matters: 404 Media obtained internal records from six companies showing Citi, Adobe and others limited employee use of flagship models and switched to cheaper versions as usage-based billing costs rose.

TechCrunch · AI

Microsoft launches Frontier Company with $2.5B to deploy enterprise AI

Microsoft formed a new operating business called Microsoft Frontier Company to deliver enterprise AI deployments using its existing AI tools. It's backed by a $2.5 billion investment and 6,000 industry and engineering experts. Commercial Business CEO Judson Althoff pushed back on the Forward-Deployed Engineer label, calling it the largest outcome-driven engineering org in the industry. The move comes two days after AWS announced a similar $1 billion internal FDE effort, following earlier moves by OpenAI and Anthropic. The post doesn't disclose specific clients or delivery timelines.

Why it matters: Microsoft put $2.5 billion and 6,000 people into Frontier Company to handle enterprise AI deployment, showing how a big vendor turns model capability into a delivery organization.

Hacker News front page

Fable and 10 other LLMs refactor a LangGraph god node, Fable's proposal ranks first

The author gave 11 LLMs a 1,500-line LangGraph god node to refactor. Fable-5's proposal scored highest in peer review, followed by GPT-5.5 and DeepSeek-4-pro. GPT-5.4 and Opus-4.7 ranked near the bottom. Each model produced full code and architecture docs, then other models cross-evaluated them. Raw data and the ranking matrix are public. Caveat: this is one refactoring task, not a general coding benchmark, but it reveals clear differences in engineering taste across models.

Why it matters: The author handed a 1,500-line LangGraph god node to 11 models to refactor and peer-review; Fable-5's plan ranked first, and the spread of approaches to the same complex code is the main point.

Hacker News front page

Training a single transformer layer can match full-parameter RL training

This paper shows that most gains from RL post-training concentrate in a few middle layers. Training a single transformer layer can recover or even exceed full-parameter RL results. The pattern holds across seven model sizes from Qwen3 and Qwen2.5, three RL algorithms (GRPO, GiGPO, Dr. GRPO), and tasks including math reasoning, code generation, and agentic decision-making. Middle layers dominate; input and output layers contribute little. Layer contribution rankings are strongly correlated across datasets, tasks, model families, and algorithms.

Why it matters: The paper proposes a layer-contribution metric for how much full-parameter RL performance a single trained layer recovers, and finds gains concentrate in the middle layers.

AI HOT (Curated Pool)

Google's AI buildout drove a 37% increase in electricity use in 2025

Google's total electricity use rose 37% YoY in 2025, driven by AI data center expansion. The company says it's balancing emissions with clean energy contracts, but the post doesn't disclose the actual clean energy coverage ratio or net emissions change. Signed contracts don't equal real-time green power delivery—grid decarbonization pace matters more.

Why it matters: Google's 2025 electricity use rose 37% year over year on AI data center growth, and the gap between clean energy contracts and actual grid decarbonization frames the numbers.

AI HOT (Curated Pool)

OpenAI reportedly offers the Trump administration a five percent stake

OpenAI is discussing giving the US government a 5% equity stake, worth over $40 billion at an $852 billion valuation. The plan would pool 5% shares from all leading US AI developers into a sovereign wealth fund modeled on the Alaska Permanent Fund, paying dividends to the government and residents. Talks have been ongoing for over a year and may require an act of Congress. Sam Altman has negotiated directly with President Trump, Commerce Secretary Lutnick, and Treasury Secretary Bessent, and recently spoke with Senator Bernie Sanders, who wants a nearly 50% public stake. Critics see the move as a way to soften political pushback and potentially pave the way for a government bailout if OpenAI's finances deteriorate. The White House did not respond to a request for comment; OpenAI declined to comment.

Why it matters: OpenAI proposed handing the US government 5% of the company, over $40 billion at an $852 billion valuation, plus a sovereign fund funded by leading AI developers, modeled on Alaska's permanent fund.

Hacker News front page

Nvidia offers startups compute power in exchange for revenue share

Nvidia plans to let fast-growing startups pay for compute with a cut of future revenue instead of cash. The program, called Nvidia Ignite, targets companies with $10M–$500M in annual revenue that aren't yet profitable. The post doesn't spell out the revenue-share percentage or deal terms—sounds like Nvidia is picking winners while GPUs are scarce, but the execution details are still vague.

Why it matters: Nvidia Ignite lets unprofitable startups with $10M to $500M in revenue offset compute costs against future revenue share, giving cash-poor teams a way to get compute without paying upfront.

The Verge · AI

OpenAI floats giving Trump administration a 5% stake in the AI boom

OpenAI proposed giving the Trump administration a 5% equity stake. Sam Altman is pitching this as a way to secure lighter regulation and avoid being treated like a public utility. The post doesn't spell out the stake's structure, voting rights, or whether the deal will actually close.

Why it matters: OpenAI pitched the Trump administration on trading a 5% company stake for lighter regulation, though the report leaves the share structure, voting rights and whether a deal happens undisclosed.

AI HOT (Curated Pool)

Qwen team's Zhu Da on consumer agents: 3× faster execution, 10× cheaper token cost vs overseas products

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.

Why it matters: Qwen's Zhu Da lays out the engineering trade-offs behind a consumer agent, with execution time down to a third and token use a tenth of comparable overseas products.

AI HOT (Curated Pool)

Unitree Robotics gets green light for STAR Market IPO registration

China's securities regulator approved Unitree Robotics' IPO registration on the STAR Market on July 1. Unitree makes quadruped and humanoid robots with fully self-developed core components, cutting key part costs to about one-third of imports. Founded in 2016 by Wang Xingxing in Hangzhou, it now has over 1,000 employees. The post doesn't disclose the fundraising amount or listing timeline; the approval is valid for 12 months.

Why it matters: Unitree has its STAR Market IPO registration approved, with 12 months to list, and its quadruped and humanoid robots use fully self-developed core parts at roughly a third of the cost of imported ones.

Hacker News front page

Experienced devs felt 20% faster with AI, measured 19% slower in controlled trial

METR ran a controlled trial with 16 experienced open-source devs across 246 tasks in familiar codebases. Measured time was ~19% slower with frontier AI tools, while self-reported speed was ~20% faster—a nearly 40-point gap. Faros AI data across 10,000+ devs shows PRs merged up 98%, PR size up >150%, review time up 91%, with roughly no net delivery change and 31% of PRs merged without review. GitClear's analysis of 200M changed lines found copy-pasted code rising and refactoring falling below 10% of changes. The bottleneck shifted from writing to reviewing, but most dashboards still track the old metric. The slowdown flips positive for juniors and greenfield work; the author frames this as a likely J-curve dip, not the final state.

Why it matters: METR ran a controlled experiment with 16 experienced developers on their own codebases: they felt about 20% faster, measured about 19% slower, a gap of nearly 40 points.

Latent Space

AIEWF Day 3: Autoresearch takes the stage, but speakers push back on full autonomy

Day 3 of AIEWF focused on autoresearch. Introspection's Roland Gavrilescu described it as an outer loop where agents maintain the system itself. Anthropic's Thariq Shihipar echoed continuous discovery in his Claude Code keynote, saying models are 'grown, not developed.' Former Google engineering lead Addy Osmani pushed back hard: the outer loop must stay human—inner loop is capability, outer loop is agency. Notion's Geoffrey Litt and Impeccable's Paul Bakaus both argued humans need to understand the code and steer the final 20%. Bakaus stated flatly there will 'never be auto.' Google's Nicole Brichtova added that cultivated expertise sees what average preference misses.

Why it matters: Day three of the AI Engineer World's Fair focused on autoresearch, which Introspection's Roland Gavrilescu described as an outer loop where agents maintain the inner loop, while several speakers held reservations about full automation.

Hacker News front page

Kimi K2.7 Code is now generally available in GitHub Copilot as the first open-weight model option

GitHub Copilot added Kimi K2.7 Code to its model picker—the first open-weight model available. It runs on Microsoft Azure and is billed at provider list pricing under usage-based billing. Rollout starts with Pro/Pro+/Max plans; Business and Enterprise admins must enable it manually. The post doesn't disclose benchmark scores, only that GitHub will monitor quality and performance.

Why it matters: Kimi K2.7 Code is now in the GitHub Copilot model picker, the first open-weight option there, with viewable and self-deployable code billed at the model vendor's published pricing.

Financial Times · Technology

OpenAI proposes handing Trump administration a 5% stake

OpenAI is considering giving the Trump administration a 5% stake as part of its restructuring into a for-profit public benefit corporation. Only the headline is available so far; the post does not disclose valuation, timeline, or how the government would hold the stake. Treat this as a negotiation signal rather than a done deal.

Why it matters: OpenAI proposed handing the Trump administration a 5% stake during its for-profit conversion, but only the headline is out: valuation, stake form and timeline are undisclosed, so treat it as a negotiating signal.

Financial Times · Technology

AI is helping workers sue their bosses—and it may break the system

The FT reports that AI tools are flooding US courts with employment lawsuits. Workers who once needed a lawyer can now use apps like DoNotPay and JusticeDirect to auto-generate complaints in minutes. EEOC charges rose 9% last year; federal labor cases jumped 10%. Judges say many AI-drafted filings cite wrong statutes or stitch together unrelated facts, clogging dockets without improving access to justice. The post doesn't provide dismissal-rate data but quotes judges warning that meritorious cases get buried in the noise.

Why it matters: Apps like DoNotPay and JusticeDirect let workers generate a lawsuit in minutes without a lawyer, and employment filings are climbing.

Hacker News front page

Snorkel releases Senior SWE-Bench, a benchmark that evaluates coding agents like senior engineers

This benchmark gives agents natural-language feature requests and bug reports instead of over-specified specs. A validation agent writes behavioral tests to check solutions, and a taste-scoring system grades whether the code fits the codebase's actual style. The dataset and scoring approach are open-sourced on GitHub.

Why it matters: Snorkel's open-source Senior SWE-Bench swaps fixed spec documents for natural-language feature requests and bug reports, with a verifier agent writing behavior tests.

AI HOT (Curated Pool)

Ant Group's AI assistant Abao opens public beta in Alipay

Ant Group's AI assistant Abao is now in public beta inside Alipay—no invite code needed. iOS and Android users can search 'Abao' to access it. Swipe right in the app to see a chat interface plus an asset page. You can speak or type a request like 'check my housing fund,' and Abao pulls up the matching mini-program and service entry, collapsing multi-step navigation into one sentence. Alipay says all fund transfers and payments still require user confirmation; Abao only runs the workflow and presents the final step. The post doesn't disclose the underlying model, the number of supported services, or future pricing.

Why it matters: Ant Group put its AI assistant Abao into public beta inside Alipay, no invite code needed, collapsing multi-mini-program errands into one conversation.

AI HOT (Curated Pool)

Emil Kowalski turns UI animation taste into Skills for AI coding agents

Emil Kowalski released three design-engineering Skills for Codex, Claude Code, and Cursor. They encode hard rules: no animation on high-frequency actions, keep UI motion under 300ms, animate only transform and opacity, start entrances from scale(0.95) + opacity:0, and respect prefers-reduced-motion. The review-animations Skill outputs a Before/After/Why table; animation-vocabulary turns vague descriptions into professional motion terms. The post doesn't say whether the Skills are open-source or paid.

Why it matters: Emil Kowalski packaged his UI animation rules into a Skill for Codex, Claude Code and Cursor, so AI coding keeps motion under 300ms and skips animation on frequent actions.

Computing Life · Share · Yage

Claude Science skips the eureka moment and starts as a compute dispatcher

Anthropic's Claude Science desktop app, announced June 30, skips the AI-scientist fantasy and targets the grunt work that eats 80% of researchers' time. It auto-pulls and cleans data from UniProt, PDB, Ensembl, and ChEMBL, then writes SLURM scripts, sets up conda environments, and submits jobs to HPC clusters—retrying on failure. By constraining the model to verifiable execution tasks, it sidesteps hallucination risks. The post does not disclose pricing or a GA date.

Why it matters: Anthropic's Claude Science desktop app, due June 30, targets the data handling that eats 80% of researchers' time rather than generating new hypotheses.

Computing Life · Share · Yage

Token pricing squeeze: buyers flee, sellers double down, builders fill the gap

Anthropic launched Claude Sonnet 5 at a promo price of $2/$10 per million tokens, rising to $3/$15 in September, while a new tokenizer adds ~30% more tokens per task—nearly doubling real cost over Sonnet 4.6. The next day Palantir's Karp called token billing “completely wrong,” arguing real value should mean outcome-based pricing. On the buy side, Uber burned its full-year AI coding budget in four months after deploying Claude Code and Cursor to 5,000 engineers, then capped per-engineer spend at $1,500/month. Microsoft cut Claude Code licenses for thousands of engineers on the last day of its fiscal year, routing them back to Copilot. GitHub Copilot's switch from flat-rate to token billing triggered a wave of cancellations as users exhausted quotas in a day or two. Sellers can't stop: OpenAI projects a $14B loss in 2026, Anthropic's monthly compute cost runs $1.25B, and Amazon shifted Anthropic payments from per-hour to per-token, pushing its own teams to distill smaller models. Builders in the middle route only 26% of requests to expensive closed-source models, add prompt caching and semantic caching to cut costs 40%–80%, and lean on open-source models whose programming traffic share has pushed Claude Opus down to 4.7%. Tokens won't disappear—they'll recede to a backend meter while front-office billing shifts to per-task, per-seat, or annual budget models.

Why it matters: Anthropic's new tokenizer adds roughly 30% more tokens for the same task, and the September promo price is expiring, so token-billed products face a quantified cost hit.

Hacker News front page

Meta Caps Internal AI Token Spending as Costs Near Billions

Meta warned ~6,000 employees that internal AI token costs are on track to hit billions this year. Employees burned through 73.7 trillion tokens in ~30 days, tracked on an internal leaderboard called 'Claudeonomics.' CTO Andrew Bosworth said token volume isn't a measure of impact. Meta is scrapping the leaderboard, rolling out an 'AI Gateway' monitoring dashboard, and will enforce formal token budgets starting in 2027. The company is also steering staff from Anthropic Claude toward its own MetaCode assistant. Uber faced a similar problem—it blew through its 2026 AI coding budget in four months and now caps spending at $1,500 per person per month.

Why it matters: Meta's internal AI token bill neared billions, so it capped usage for about 6,000 staff, with the CTO noting usage is not output.

Hacker News front page

Zhipu launches ZCode 3.0, a coding agent deeply integrated with GLM-5.2

Zhipu released ZCode 3.0, a Claude Code-like coding agent deeply integrated with its GLM-5.2 model. It breaks complex tasks into multi-step goals and supports remote bot triggers via WeChat or Telegram. Pricing starts at $16.2/mo for Lite, $64.8/mo for Pro, and $144/mo for Max. The post does not disclose context window size, benchmark scores, or direct comparisons with Copilot or Cursor.

Why it matters: Zhipu's ZCode 3.0 runs on its own GLM-5.2 model, sits close to Claude Code, and takes remote tasks via WeChat or Telegram.

Latent Space

How Cursor's Forward Deployed Engineers build AI software factories in the enterprise

Cursor VP Pauline Brunet explained at AIEWF how her Forward Deployed Engineers embed Cursor's agents across the full software lifecycle—planning, coding, testing, and deployment—to build an 'AI software factory.' The team hires engineers with 5+ years of experience and plans to grow 10x by year-end. The main enterprise bottleneck: individual early adopters are productive, but scaling long-running agents across teams requires top-down leadership commitment.

Why it matters: Cursor's VP Pauline Brunet described forward-deployed engineers who embed in customer systems and workflows to roll out its coding assistants and agents.

TechCrunch · AI

Cloudflare sets a Sept 15 deadline for AI companies to split search crawlers from AI crawlers or face default blocks

Cloudflare will start blocking mixed-use crawlers by default on ad-supported pages from September 15, 2026. The block applies to new customers, new sites from existing customers, and all free-tier users. Site owners can override it, but the default stance is: separate search crawling from AI training and agent use, or get blocked. The post doesn't say how many sites this covers or whether Cloudflare plans its own paid licensing path.

Why it matters: Cloudflare will block mixed-use crawlers by default on ad-bearing pages from September 15, 2026, covering new customers, new sites for existing ones and free users, pushing the AI training payment fight into infrastructure defaults.

Hacker News front page

Over one-third of ChatGPT conversations involve fiction generation

Neel Gupta and co-authors analyzed over 500,000 anonymized ChatGPT conversations and found more than one-third involve fiction generation—original stories, roleplay, fanfiction, and erotica. Power users dominate this activity; some repeatedly request and revise the same narratives, which the authors call 'infinite story demanders.' Users gravitate toward fanfiction, erotica, generic forms, immediacy, and niche story-element combos. The paper introduces the concept of a 'solipsistic reader-writer' who produces and consumes fiction inside a closed loop with a machine. The post does not disclose the time span or geographic breakdown of the 500,000 conversations.

Why it matters: A study of over 500,000 English ChatGPT conversations found more than a third went to generating fiction, a share driven mainly by heavy users rewriting the same story.

Jul 1Wednesday

MIT Technology Review · AI

LLMs are stuck in a groupthink groove. This startup is trying to get them out.

Mainstream LLMs converge on near-identical answers to open-ended prompts—ask ChatGPT or Claude for a random number and you'll almost always get 7. Australian startup Springboards built Flint, a model trained to produce wider response variety, like returning 3.7916 for that same prompt. A NeurIPS 2025 best paper found 25 different models mostly repeat 'time is a river' when asked for a metaphor. Springboards calls this 'lost information' and targets creative professionals who need to break out of the groupthink rut.

Why it matters: Australian startup Springboards says mainstream models converge on open questions, and a NeurIPS 2025 best paper found 25 models mostly wrote the same time metaphor.

Latent Space

Warp CEO Zach Lloyd on why software factories are the next phase of coding

At AI Engineer World’s Fair, Warp CEO Zach Lloyd argued coding is shifting from interactive agent use to fully automated development loops. He calls this a 'software factory'—agents continuously triage, implement, review, verify, ship, and monitor changes. Warp’s new platform Oz lets teams set up such factories, plugging into Jira, Slack, and GitHub, with configurable human checkpoints. Lloyd expects most major projects to adopt some form of automated factory within a year. Warp open-sourced its terminal tool in April and is now pivoting hard toward agent orchestration.

Why it matters: Warp CEO Zach Lloyd argues coding is moving from people directing single AI agents to fully automated pipelines that triage, write, review, verify, ship and monitor.

TechCrunch · AI

Venice AI hits unicorn status with $65M Series A for its privacy-first, no-log platform

Venice AI raised $65M at a $1B+ valuation. It sells a no-log, uncensored chatbot and is already profitable with a $70M+ annualized run rate. CEO Erik Voorhees says users will pay for privacy. The post doesn't disclose which models it runs or its inference cost breakdown.

Why it matters: Venice AI raised a $65M Series A at over $1B valuation on privacy and no-censorship, with annualized revenue past $70M, a rare paid privacy case in the app layer.

TechCrunch · AI

Google's agentic assistant Gemini Spark is now on Mac

Google brought Gemini Spark, its AI agent for file sorting and cross-app tasks, to Mac. It can read local files—turning invoices into a budget sheet, for example—and will later support remote phone-to-desktop commands. It's in beta, only for Google One AI Premium subscribers.

Why it matters: Google's agentic assistant reaches Mac, where it can read local files and act across apps, turning invoices into budget spreadsheets, in beta for Google One AI Premium subscribers.

AI HOT (Curated Pool)

Meta plans to sell excess AI compute, following SpaceX's playbook

Meta is building a cloud infrastructure business to sell AI compute and model access, putting it in direct competition with AWS, Google Cloud, and Microsoft Azure. The move comes weeks after SpaceX leased its Colossus 1 data center capacity to Anthropic. The pattern suggests data center owners, not model builders, may end up winning the AI race.

Why it matters: Meta is preparing a cloud infrastructure business, selling AI compute and model access against AWS, Google Cloud and Microsoft Azure, adding a new contender to cloud competition.

Up to 50 pages are available; use search or topics for older items.