Skip to content

#编码

10 today

Apr 27Monday

QbitAI · WeChat

DeepSeek V4 Cuts Prices Permanently; Cached Inputs Get 90% Off, Coding Test Costs Drop 83%

DeepSeek V4 cut prices twice in two days: input/output pricing is 75% lower, with cached inputs getting another 90% off. QbitAI’s coding test fell from 31.73 yuan for 35M tokens to 5.34 yuan under new pricing, an 83% drop. The key case is high cache-hit workloads, with V4-Pro at about 95–96% cache hits.

Why it matters: DeepSeek V4 cut input and output prices 75% twice in two days, with cached input down another 90%, and QbitAI measured coding cost falling from 31.73 yuan to 5.34 yuan.

Synced · WeChat

ACL 2026: Sending AI “~” May Cause It to Delete Your Home Directory

ACL 2026 accepted an LLM safety paper on emoticon semantic confusion. The team tested 6 models with 3,757 cases; average confusion was 38.6%, with over 90% silent failures. The key risk is agent execution, where “ignore emoticons” prompts had limited effect.

Why it matters: A 3,757-sample test across six models found emoji tripped them up 38.6% of the time, and over 90% of those errors passed silently.

OpenAI News

An Open-Source Spec for Orchestration: Symphony

OpenAI released Symphony, an open-source spec for Codex orchestration. The RSS snippet says it turns issue trackers into always-on agent systems; the post does not disclose spec details, license, APIs, or benchmarks.

Why it matters: OpenAI open-sourced the Symphony orchestration spec, turning a Linear board into a dispatch point for coding agents where humans only review results; internal teams merged five times as many PRs in three weeks.

Hacker News front page

If You Stop Hiring Juniors, Your Senior Engineers Own You

Justin Smestad argues that firms stopping junior hiring in 2026 risk costly senior-heavy teams by 2030. The mechanism: a senior can demand a 40% raise; without a two-year bench, replacement may take six months. The key issue is pipeline leverage, not quarterly headcount savings.

Why it matters: Justin Smestad argues that halting junior hiring from 2026 leaves a 2030 team of expensive veterans with no bench, letting seniors demand 40% raises.

Apr 26Sunday

Hacker News front page

Why SWE-bench Verified No Longer Measures Frontier Coding Capabilities

OpenAI stopped reporting SWE-bench Verified scores and recommends SWE-bench Pro instead. It audited 138 tasks that o3 failed inconsistently across 64 runs and found 59.4% had test or prompt flaws. The key issue is contamination: tested frontier models reproduced some gold patches or task details.

Why it matters: OpenAI audited 138 problems o3 repeatedly failed and found at least 59.4% had overly rigid test cases or vague descriptions, which explains why it retired the benchmark.

Hacker News front page

The West Forgot How to Make Things. Now It's Forgetting How to Code

Denis Stetskov compares AI coding to 7 defense knowledge-loss cases: a 2022 Stinger order delivers in 2026. The post cites EU shell capacity at 230,000/year and a 1M-shell pledge met 9 months late; the risk is the junior engineer pipeline, not single-task coding speed.

Why it matters: Uses the West's slow Stinger missile and EU shell production as a parallel for what AI coding does to software.

Hacker News front page

Simulacrum of Knowledge Work

The author argued on 2026-04-25 that LLMs break surface-quality proxies in knowledge work. Examples include market reports and code review, ending in skims, LGTM, and a 17th Claude Code session. The critique targets evaluation: corpus likelihood or RLHF preference, not truth.

Why it matters: The author argues LLMs make surface polish useless as a proxy for quality in knowledge work, using a market analysis report as the example.

Hacker News front page

Using Coding Assistance Tools to Revive Projects You Never Were Going to Finish

Matthew Brunelle used Claude Code with Opus 4.6 to rebuild a YouTube Music-to-OpenSubsonic connector, listing 6 setup steps. The stack used FastAPI, Pydantic, ytmusicapi, and yt-dlp, with Feishin logs used to fix .view suffix handling. The useful point: a clear spec plus human review beat one-shot generation.

Why it matters: The author used Claude Code to rebuild a shelved project into a YouTube Music to OpenSubsonic bridge, letting players like Feishin stream YouTube songs, with a spec-first, human-reviewed workflow worth copying.

Apr 25Saturday

Latent Space

DeepSeek V4 Pro and Flash released, runnable on Huawei Ascend chips

DeepSeek released V4 Pro and V4 Flash, with 1.6T/49B active and 284B/13B active parameters. Both support 1M-token context, Base/Instruct variants, and an MIT license; the report claims 27% FLOPs and 10% KV cache versus V3.2 at 1M tokens. The key point is Huawei CANN compatibility, not just benchmarks, because it reduces CUDA dependence.

Why it matters: DeepSeek's V4 Pro and V4 Flash both take a 1M-token context, cut long-text compute to 27% of V3.2, and run on Huawei Ascend chips.

Computing Life · Share · Yage

TPU vs. CUDA: A Post-Cloud Next 2026 Assessment

Google announced TPU 8t/8i, TorchTPU, and an Anthropic deal at Cloud Next 2026; TPU 8i is slated for H2 2027 volume production. 8i has 288GB HBM, 8.6TB/s bandwidth, and 384MB SRAM; TorchTPU runs PyTorch on TPU, but the post says independent benchmarks are missing. The key crack is vLLM inference, while the author says TPU will not replace NVIDIA within 18-24 months.

Why it matters: At Cloud Next 2026 Google split its eighth-gen TPU into 8t for training and 8i for inference, and launched TorchTPU so PyTorch code runs directly without the old bridge layer.

MIT Technology Review · AI

Three reasons why DeepSeek’s new model matters

DeepSeek released a V4 preview with two versions: V4-Pro and V4-Flash. V4-Pro costs $1.74/M input tokens and $3.48/M output tokens; V4-Flash is about $0.14/$0.28, and both support 1M-token context. The key point is attention efficiency and open weights pressuring agentic coding costs.

Why it matters: DeepSeek's V4 preview prices Pro and Flash about twelve times apart, both with 1M-token context, so readers can compare open-model costs for long text.

Hacker News front page

Google Flow Music

Google Flow Music launched a web creation entry with six sections: songs, playlists, Spaces, videos, projects, and Turntable. The page says Producer creates full songs with Lyria 3, and AI music videos use Veo. Pricing, regions, model specs, and rights terms are not disclosed.

Why it matters: Google launched the web version of Flow Music, putting Lyria 3 song generation and Veo music videos behind one entry point, with the conversational Producer tool making full tracks with vocals.

Apr 24Friday

Hacker News front page

Affirm Retooled Its Engineering Organization for Agentic Software Development in One Week

In February 2026, Affirm paused normal engineering work for one week and asked 800+ engineers to complete a full agentic workflow from ideation to submitted PR; it says over 60% of PRs are now agent-assisted. The post adds that 80%+ of engineers were weekly active users of AI dev tools by December 2025, and a nine-engineer group spent two weeks defining a default workflow around Claude Code, local-first development, and human checkpoints; the captured body does not fully disclose later implementation details or measured outcomes.

Why it matters: Affirm paused normal development for a week so 800-plus engineers could ship real requests to PRs with coding agents, and over 60% of the resulting PRs involved an agent.

The Verge · AI

China’s DeepSeek previews new AI model a year after jolling US rivals

DeepSeek released a preview of its open-source V4 model on Friday and said it can compete with closed systems from Anthropic, Google, and OpenAI. The RSS snippet says V4 improves coding and highlights compatibility with Huawei tech; parameter count, benchmark scores, and rollout details are not disclosed. The part to watch is the pairing of agent-focused coding gains with tighter alignment to China’s domestic chip stack.

Why it matters: DeepSeek released an open-source V4 preview, claiming better coding and compatibility with Huawei hardware, which ties model progress to China's chip stack.

Synced · WeChat

Anthropic confirms three bugs caused Claude Code's apparent quality drop

Anthropic said Claude Code's quality drop over the past month came from 3 harness and prompt issues, while model capability itself and the Claude API were unchanged. The issues were a Mar. 4 default reasoning shift from high to medium, a Mar. 26 session-cache bug, and an Apr. 16 25/100-word prompt limit; fixes or rollbacks landed on Apr. 7, Apr. 10, and Apr. 20.

Why it matters: Anthropic traced a month of degraded Claude Code experience to three engineering bugs, including a lowered default reasoning effort and a session cache fault, clearing the model itself.

QbitAI · WeChat

Claude admits three issues: downgraded reasoning, cleared memory, and constrained output

Anthropic said on April 23 that three Claude issues hurt quality: Claude Code default reasoning was changed from high to medium on March 4 while the UI still showed high. A March 26 cache bug cleared thinking state every turn for 15 days, and an April 16 prompt limit of 25 words between tool calls and 100 words in final replies cut Opus 4.6/4.7 by 3% before a rollback four days later.

Why it matters: Anthropic publicly reviewed three recent Claude quality problems, including a default reasoning effort lowered from high to medium while still displaying high, a gap between what users pay for and what they get.

Latent Space

GPT 5.5 and OpenAI Codex Superapp

OpenAI launched GPT-5.5 for ChatGPT and Codex, while API access is delayed for safeguards. The post cites 82.7% Terminal-Bench 2.0, 58.6% SWE-Bench Pro, and a 1M API context window. The sharper signal is Codex: browser control and Prism integration point to a desktop superapp strategy.

Why it matters: OpenAI shipped GPT-5.5 in ChatGPT and Codex first, with API access delayed for safety hardening, scoring 82.7% on Terminal-Bench 2.0 and 58.6% on SWE-Bench Pro.

X · @op7418

DeepSeek V4 arrives with Flash and Pro variants

DeepSeek released V4 with two variants, Flash and Pro. The RSS snippet says it supports JSON output, tool calling, dialogue prefix continuation, and FIM completion; Flash costs ¥0.2/¥1 per million input/output tokens, while Pro costs ¥1/¥12. At 1M context, output pricing doubles.

Why it matters: DeepSeek's V4 ships in Flash and Pro tiers, with Flash at 0.2 yuan per million input tokens and 1 yuan output, Pro at 1 yuan input and 12 yuan output, doubling at a million-token context.

Bloomberg Technology

AI Coding Firm Cognition in Funding Talks at $25 Billion Value

Cognition is in early talks to raise funding at a $25 billion valuation, more than double its prior valuation. The RSS snippet says demand for AI software-development firms is rising, but the post does not disclose investors, round size, or timing.

Why it matters: Cognition is negotiating a new round at a reported $25 billion valuation, more than double its last round, a pace that shows continued market bets on AI coding.

X · @dotey

Codex now supports GPT-5.5 and adds five capability upgrades

Codex now supports GPT-5.5 and adds 5 upgrades aimed at moving it from a coding tool to an agent that can execute longer tasks. The RSS snippet says it can control browsers and computers, create files in Microsoft Office and Google Drive, and use gpt-image-2; an auto-review mode invokes a separate review agent for high-risk actions. What matters is longer task chains, but the post does not disclose pricing, rollout scope, or safety thresholds.

Why it matters: Codex adds GPT-5.5 plus browser control, Office and Google Drive document generation and auto review, moving from a coding tool to an agent that runs multi-step tasks.

X · @dotey

OpenAI launches GPT-5.5 for paid ChatGPT and enterprise users, with Codex; API coming soon

OpenAI launched GPT-5.5 for ChatGPT Plus, Pro, Business, and Enterprise users, alongside Codex. OpenAI says per-token latency matches GPT-5.4, while Terminal-Bench 2.0 rises to 82.7% from 75.1%; API pricing is $5 per 1M input tokens and $30 per 1M output tokens with a 1M-token context. The key detail is efficiency: the post says GPT-5.5 uses about half the total tokens of frontier rival coding models at the same intelligence level.

Why it matters: OpenAI opened GPT-5.5 to ChatGPT paid users and Codex, saying per-token latency matches GPT-5.4 while the same tasks use fewer tokens.

Hacker News front page

GPT-5.5: Mythos-Like Hacking, Open to All

XBOW says GPT-5.5 cut miss rate to 10% on its real-vulnerability benchmark, versus 40% for GPT-5 and 18% for Opus 4.6. It scored 97.5% on visual acuity and used about half the login iterations of the next-best model. The key point is black-box testing: GPT-5.5 without source beat GPT-5 with source.

Why it matters: Security firm XBOW tested GPT-5.5 against a real vulnerability set, and its black-box miss rate fell to 10%, below GPT-5 white-box at 40% and Opus 4.6 at 18%.

The Verge · AI

OpenAI says its new GPT-5.5 model is more efficient and better at coding

OpenAI announced GPT-5.5 and says it is more efficient and stronger at coding than GPT-5.4, which shipped last month. The RSS snippet says it handles coding, debugging, online research, and cross-tool work on spreadsheets and documents; the post does not disclose pricing, context window, or benchmark scores.

Why it matters: OpenAI's GPT-5.5 leans on coding and debugging and is claimed to be more efficient than GPT-5.4, with no benchmarks, context window or pricing disclosed.

Hacker News front page

An update on recent Claude Code quality reports

Anthropic said three product-layer changes degraded Claude Code quality for Sonnet 4.6, Opus 4.6, and Opus 4.7, while the API was unaffected; all were fixed on April 20 in v2.1.116. The changes were lowering default reasoning effort on March 4, a March 26 bug that cleared prior thinking every turn after sessions sat idle for over an hour, and an April 16 prompt tweak to reduce verbosity that hurt coding quality. The signal for practitioners is sharp: product and prompt changes can degrade code performance even when model and inference evals do not reproduce it early.

Why it matters: Anthropic traced a Claude Code quality drop to three product-layer changes, not the API or model, showing non-model tweaks can shape the coding experience.

Apr 23Thursday

QbitAI · WeChat

Qwen3.6-27B open-weights, beats its 397B flagship predecessor on agentic coding

Qwen released Qwen3.6-27B and says it beats Qwen3.5-397B on 4 agentic coding benchmarks with about 1/15 the parameters. The post cites SkillsBench rising from 30.0 to 48.2, GPQA Diamond at 87.8, and AIME26 at 94.1; it uses a dense architecture, Thinking Preservation, and Gated DeltaNet, with weights on Hugging Face and ModelScope.

Why it matters: Qwen released Qwen3.6-27B, a 27B dense model it says beats its own 397B previous flagship on four agentic coding benchmarks.

OpenAI News

Introducing GPT-5.5

OpenAI introduced GPT-5.5 and says it targets complex cross-tool tasks such as coding, research, and data analysis. The RSS snippet only confirms “faster” and “more capable”; the post does not disclose benchmarks, context window, pricing, release timing, or availability, which are the details practitioners should watch.

Why it matters: OpenAI's GPT-5.5 is pitched at multi-step work, planning and calling tools on its own, with an 82.7% score on Terminal-Bench 2.0 and fewer tokens per task.

Xinzhiyuan · WeChat

Historic moment: Anthropic nears $1 trillion on private secondary markets, surpassing OpenAI for the first time

Anthropic was quoted at $1.05T-$1.15T on private secondary markets, above OpenAI’s roughly $880B quotes on similar platforms. The post attributes the rerating to scarce float, a sharp rise from a $380B funding valuation three months earlier, and momentum around Claude Code and revenue growth; it does not disclose trade volume, revenue figures, or company confirmation. Do not confuse this with a new funding valuation: these are secondary-market quotes on platforms such as Forge Global.

Why it matters: Anthropic's secondary-market quotes of $1.05tn to $1.15tn now top OpenAI's roughly $880bn, three months after a $380bn funding valuation, exposing the gap between the two numbers.

New York Times Chinese

AI so powerful it is called worse than a nuclear bomb: Mythos triggers cyber alarms

Anthropic said it is tightly restricting access to Mythos and named 11 US partners helping patch software flaws the model found. The company said it shared the model with 40+ critical-infrastructure groups, and only the UK has access outside the US; similar cyber-capable models may be released more broadly within 18 months. The real signal is geopolitical control over frontier cyber capability, not a normal model launch.

Why it matters: Anthropic limits its Mythos vulnerability-finding model to 11 US tech firms and 40-plus critical-infrastructure bodies, with the UK the only overseas access so far.

Bloomberg Technology

Andreessen, Thrive Poised for Windfall From SpaceX's Bid for Cursor

If SpaceX acquires AI coding startup Cursor for $60 billion, Andreessen Horowitz and Thrive Capital stand to gain billions. The RSS snippet says Andreessen holds about 10%, worth roughly $6 billion at that price; the post does not disclose whether a deal is signed or Thrive's exact stake.

Why it matters: SpaceX is weighing a $60 billion buyout of Cursor, and a16z's roughly 10% stake would be worth about $6 billion if it closes.

Latent Space

Shopify’s AI Phase Transition: 2026 Usage Explosion, Unlimited Opus-4.6 Budget, Tangle

Shopify CTO Mikhail Parakhin detailed its AI stack across 3 projects: Tangle, Tangent, and SimGym. The post says Shopify is a 20-year, $200B software company, but does not disclose exact 2026 usage figures. The key shift is from code generation to review, CI/CD, and deployment stability.

Why it matters: Shopify's CTO walks through how internal AI tool use shifted, with usage spiking after a December 2025 jump in model quality and CLI tools growing faster than IDE plugins.

Hacker News front page

Coding Models Are Doing Too Much

The author programmatically corrupts 400 BigCodeBench problems with single-point bugs to test whether coding models over-edit code during fixes. The post defines the minimal fix as exactly reversing the corruption and measures excess changes with token-level Python Levenshtein distance. The provided body does not disclose final results, model rankings, or training gains.

Why it matters: The author plants single-point bugs in 400 BigCodeBench tasks and uses token-level edit distance to measure how much models over-rewrite when fixing bugs, then tries to train them to make minimal edits.

Hacker News front page

Introducing Parallel Agents in Zed

Zed released Parallel Agents on April 22, 2026, letting multiple agents run in parallel in one window. The new Threads Sidebar sets per-thread folder and repo access, and supports stop, archive, and new-thread actions; the new default layout is opt-in for existing users. The key detail is permission scoping and thread orchestration, not just “multiple agents.”

Why it matters: Zed's April 22 update runs several AI agents side by side in one window, with per-thread folder and repo access limits.

Hacker News front page

Startups Brag They Spend More Money on AI Than Human Employees

Swan AI CEO Amos Bar-Joseph said his 4-person startup spent $113,000 on Claude in one month and treated that bill as headcount budget spent on AI instead of hires. The post says Swan targets $10M ARR with fewer than 10 people and cites Fundable AI claiming AI can replace a 15-person document team; the real signal is that token spend is being used as a growth metric, not proven ROI.

Why it matters: A four-person team spending $113,000 a month on Claude shows what replacing headcount with model calls actually costs.

Apr 22Wednesday

Hacker News front page

Show HN submissions tripled and are now mostly vibe-coded

Adrian Krebs scored 500 recent Show HN landing pages and says submissions have tripled, with 67% of pages triggering at least 2 AI design patterns. The method used Playwright plus an in-page script to check DOM and computed styles across 15 deterministic CSS/DOM signals; manual QA found about 5% to 10% false positives. The real signal is not model quality, but fast homogenization from AI default frontend templates.

Why it matters: Adrian Krebs scripted a scan of 500 Show HN landing pages and found 67% hit at least two AI design tells, giving a measured read on how alike AI-generated front ends have become.

Hacker News front page

Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model

Qwen released the open-weight 27B dense model Qwen3.6-27B and made it available in Qwen Studio. It scores 77.2 on SWE-bench Verified vs. 76.2 for Qwen3.5-397B-A17B, and 59.3 on Terminal-Bench 2.0 under a 256K context and 3-hour timeout. The real takeaway is deployment: this is not a larger MoE, but a denser 27B model with stronger coding results.

Why it matters: Qwen open-sourced Qwen3.6-27B, a 27B dense coding model that scores 77.2 on SWE-bench Verified, beating the previous 397B-parameter MoE flagship, making the dense-versus-MoE comparison the draw.

OpenAI News

Introducing workspace agents in ChatGPT

OpenAI introduced workspace agents in ChatGPT, describing them as Codex-powered agents that automate complex workflows in the cloud. The RSS snippet confirms secure work across tools for teams, but the post does not disclose pricing, availability, supported tools, or performance metrics.

Why it matters: OpenAI is putting Codex-powered workspace agents into paid ChatGPT plans, so team automation moves from chat to running multi-step tasks across tools in the cloud.

The Verge · AI

Anthropic’s most dangerous AI model just fell into the wrong hands

Anthropic’s Claude Mythos Preview was accessed by a small group of unauthorized users through a contractor’s access plus common internet sleuthing tools. The snippet says the model can identify and exploit flaws in major operating systems and browsers; the post does not disclose the group size, dwell time, or remediation status. The key issue is access control failure, not the headline’s danger framing.

Why it matters: A contractor permission slip let Discord users reach Anthropic's internal Claude Mythos Preview for two weeks, showing how thin third-party access control is at frontier labs.

Xinzhiyuan · WeChat

Musk sets Cursor deal terms: SpaceX can buy it for $60B or pay a $10B collaboration fee

SpaceX disclosed terms to work with Cursor: it can acquire the startup this year for $60B or pay a $10B collaboration fee. The post says xAI already supplies Colossus compute to train Cursor's Composer, while Cursor had over $2B annualized revenue and 1M+ daily users by Feb. 2026. The key point is the bundle: SpaceX gets a coding product, and Cursor gets model and compute support.

Why it matters: SpaceX gave Cursor a take-it-or-leave-it choice this year, and the two price tags shape how the AI coding race gets financed.

Financial Times · Technology

SpaceX obtains right to buy AI start-up Cursor for $60bn

The headline says SpaceX obtained the right to buy AI start-up Cursor for $60bn. The RSS snippet adds one fact: Elon Musk’s rocket and AI group is trying to catch up with OpenAI and Anthropic; the post does not disclose the trigger, timeline, deal structure, or regulatory terms. The key issue is whether this is tied to future financing or control rights, not a completed acquisition.

Why it matters: SpaceX holds the right to buy AI coding tool Cursor for $60bn, a figure that frames how code assistants are valued as Musk's companies chase OpenAI and Anthropic.

X · @dotey

Anthropic quietly removed Claude Code from the $20 Pro plan on its pricing page without an announcement

Anthropic was spotted removing Claude Code from the $20 Pro plan on its pricing comparison page without an announcement. The snippet says help docs also removed the inclusion, while the Claude Code product page and support bot still say it is included, and some Pro users report access still works; the post does not disclose Anthropic’s formal explanation or effective date. The key issue is price floor: if confirmed, entry cost for Claude Code rises from $20 to $100 per month.

Why it matters: Anthropic's own pages disagree on whether Pro still includes Claude Code, and no explanation or effective date has been given.