Skip to content

OpenAI / ChatGPT

Everything OpenAI: the GPT models, ChatGPT and Sora, company strategy and people moves.

Latest picks

1321–1340 of 1,724

May 1Friday

Xinzhiyuan · WeChat

Claude Code's Real Story: 98.4% of What Works Is Engineering, Not AI

VILA-Lab analyzed 512,000 lines of Claude Code v2.1.88 and found 1.6% tied to AI decision logic. The other 98.4% is deterministic infrastructure: permissions, context, tool routing, and error recovery. The key shift is harness design, not longer prompts.

Why it matters: VILA-Lab reverse-engineered Claude Code's 512,000 lines and found only 1.6% relates to AI decisions; the other 98.4% is engineering infrastructure like permission control, context management and tool routing.

Xinzhiyuan · WeChat

OpenAI upgrades Codex to control Macs and run cross-app tasks

OpenAI upgraded Codex with Slack, Google Workspace, and Microsoft 365 integrations. Mike Russell tested Codex on a Mac across Adobe Audition, Photoshop, and Firefly, finishing in about 8 minutes with an 85–90 score. The key shift is OS-level computer control, not code completion.

Why it matters: OpenAI wired Codex into Slack, Google Workspace and Microsoft 365, letting the assistant act across apps on a Mac, shifting the focus from writing code to doing everyday office tasks.

Synced · WeChat

Researchers Estimate GPT, Claude, and Gemini Parameter Counts Using API Calls

Bojie Li posted IKP on arXiv to estimate parameter counts of 188 LLMs from 27 vendors via black-box API calls. The dataset has 1,400 questions across 7 rarity tiers, fitted on 89 open models with R²=0.917. Debate centers on synthetic data, MoE effects, and a 90% interval of 0.3x to 3x.

Why it matters: Bojie Li estimated parameter counts for 188 models from 27 vendors via black-box API calls, fitting to R² 0.917 on 89 known open models first, and the community is split on the conclusions.

Latent Space

[AINews] Agents for Everything Else: Codex for Knowledge Work, Claude for Creative Work

OpenAI expanded Codex to non-coding work, with CUA reported 42% faster. The update connects Microsoft, Google, and Salesforce, covering docs, slides, spreadsheets, research, and planning. The key signal is GUI-agent productization, not one benchmark score.

Why it matters: OpenAI pushed Codex beyond code into documents, slides and spreadsheets with Microsoft, Google and Salesforce integrations, while Anthropic added Blender and Adobe creative apps to Claude.

Hacker News front page

Show HN: Pu.sh – a full coding-agent harness in 400 lines of shell

Pu.sh ships a coding-agent harness in about 400 lines of shell, using only sh, curl, and awk. It supports Anthropic and OpenAI, 7 tools, REPL, auto-compaction, checkpoint/resume, pipe mode, and 90 no-API tests. It excludes TUI, streaming, images, OAuth, and Windows.

Why it matters: Pu.sh builds a full coding-agent harness in about 400 lines of shell, covering tool calls, context compaction and resumable runs, a minimal readable reference for anyone who wants the core agent loop without heavy dependencies.

TechCrunch · AI

After Dissing Anthropic for Limiting Mythos, OpenAI Restricts Access to Cyber, Too

OpenAI will first roll out GPT-5.5 Cyber only to “critical cyber defenders.” The RSS snippet does not disclose eligibility rules, pricing, or launch timing. The access-tiering model is the key detail for practitioners.

Why it matters: OpenAI gave GPT-5.5 Cyber first to critical cyber defenders while ordinary users wait, after Sam Altman criticized Anthropic for limiting Mythos.

The Verge · AI

Elon Musk confirms xAI used OpenAI’s models to train Grok

Elon Musk testified Thursday in a California federal court that xAI used OpenAI models to improve Grok. The mechanism described is model distillation: a larger teacher model transfers knowledge to a smaller student model. The post does not disclose which OpenAI models, data volume, or training runs.

Why it matters: Musk testified in California federal court that xAI partly used OpenAI models to improve Grok through distillation, laying out how the two firms' models relate.

TechCrunch · AI

Elon Musk testifies that xAI trained Grok on OpenAI models

Elon Musk testified that xAI trained Grok on OpenAI models. The post only says distillation concerns frontier labs; it does not disclose scale, model versions, or case context.

Why it matters: Musk's testimony in California federal court that xAI trained Grok partly on OpenAI models moves distillation from rumor into the court record.

Bloomberg Technology

Anthropic Weighs Funding Offers at Over $900 Billion Valuation

Anthropic is weighing a new funding round at a valuation above $900 billion. Bloomberg cites sources; if completed, Anthropic would overtake OpenAI by valuation. The post does not disclose round size, investors, or timing.

Why it matters: A reported round above $900 billion would put Anthropic ahead of OpenAI as the most valuable AI startup.

The Verge · AI

How the new Microsoft and OpenAI deal breaks down

Microsoft updated its OpenAI deal Monday, letting OpenAI offer products and services across all cloud providers. The RSS snippet does not disclose full contract terms, revenue sharing, or compute commitments. The key shift is the end of cloud exclusivity.

Why it matters: OpenAI can now sell to any cloud provider, ending Microsoft's exclusive hold on its cloud business.

Apr 30Thursday

The Verge · AI

OpenAI talks about not talking about goblins

OpenAI explained instructions telling its coding model to avoid goblins and similar creatures after Wired reported them. OpenAI says GPT-5.1’s “Nerdy” personality began using creature metaphors; the post does not disclose the full fix.

Why it matters: OpenAI says a new GPT-5.1 personality that leans on fantasy metaphors like goblins bled into other models, and it doesn't detail the fix.

Xinzhiyuan · WeChat

AI Raw Proofs Pile Up on GitHub as Terence Tao Says Solving Alone Is Not Enough

Terence Tao says math is shifting from proof scarcity to proof abundance, with 20-plus AI solutions pending assessment on an Erdős problems GitHub page. The post says GPT-5.4 Pro generated an Erdős #1196 approach in 80 minutes, and Tao verified the core within 24 hours. The key issue is verification and digestion workflow, not raw proof count.

Why it matters: Terence Tao says math's bottleneck has flipped from too few solutions to too many; one Erdős problem page has 20-plus AI-generated solutions waiting for review, leaving the field to work out verification and digestion.

QbitAI · WeChat

OpenAI Explains Why GPT-5.5 Keeps Saying “Goblin”

OpenAI says GPT-5.5’s “goblin” habit came from Nerd-persona rewards and training transfer. After GPT-5.1, ChatGPT’s “goblin” use rose 175%; Nerd replies were 2.5% of all replies but 66.7% of goblin mentions. The key issue is reward bias spreading through RL, rollouts, and SFT.

Why it matters: OpenAI traced why GPT-5.5 kept saying goblin, showing the reward signal behind a nerd persona spread through reinforcement learning, lifting use of the word 175% even though such replies were only 2.5% of output.

Bloomberg Technology

SoftBank’s $40 Billion Loan for OpenAI Stake Draws More Banks

SoftBank signed a $40 billion bridge loan for its OpenAI investment, and banks brought more lenders into syndication. The post cites people familiar with the matter but does not disclose pricing, maturity, collateral, or OpenAI stake size.

Why it matters: SoftBank signed a $40 billion bridge loan for its OpenAI stake, and lead banks are splitting it among more lenders, showing how the syndication is progressing.

Latent Space

[AINews] The Inference Inflection

Latent Space argues inference demand has hit an inflection point, citing its Apr 28-29, 2026 AINews roundup. Jensen Huang is quoted saying per-task compute rose about 10,000x in two years, with usage up about 100x. The key watchpoints are CPU sandboxes, agent harnesses, and split inference workloads.

Why it matters: Jensen Huang's GTC numbers show per-task compute up roughly 10,000x in two years and total usage up about 100x, and Latent Space ties that inference demand to real workloads like CPU sandboxes and agent scheduling.

TechCrunch · AI

Satya Nadella says he’s ready to ‘exploit’ the new OpenAI deal

Satya Nadella said Microsoft will use the new OpenAI deal to offer OpenAI tech to cloud customers without paying. The post only discloses Nadella’s quote and the access mechanism, not term length, pricing, or product scope.

Why it matters: Nadella told analysts Microsoft no longer pays for OpenAI models and agents it resells to cloud customers, and his word for the license was "exploit."

Bloomberg Technology

Anthropic Considering Funding Offers at Over $900 Billion Value

Anthropic is weighing a new funding round at a valuation above $900 billion. The post cites people familiar with the matter but does not disclose round size, investors, or timing. The key signal is the valuation anchor versus OpenAI.

Why it matters: Anthropic is weighing a new funding offer at a valuation above $900B, putting it in OpenAI's range and showing where the top AI valuation benchmark now sits.

OpenAI News

Where the goblins came from

OpenAI posted about goblin outputs in GPT-5; only an RSS snippet is available. The snippet names timeline, root cause, and fixes, but does not disclose mechanisms or conditions. The key issue is how personality-driven quirks enter model behavior.

Why it matters: OpenAI explains where GPT-5.1's goblin metaphors come from: a reward model trained to score nerd style favored fantasy-creature words, giving higher scores in 76.2% of the dataset.

The Verge · AI

All the Evidence Unveiled So Far in Musk v. Altman

The Verge summarizes Musk v. Altman trial exhibits, including emails, photos, and corporate documents. The snippet says Jensen Huang gave OpenAI a scarce supercomputer and Musk shaped its mission; the post does not disclose the full exhibit list or trial schedule.

Why it matters: The Verge gathered the emails, photos and company documents made public so far in Musk v. Altman, reconstructing OpenAI's early internal operations from 2015 on, including Jensen Huang supplying scarce supercomputing.

Hacker News front page

Show HN: A New Benchmark for Testing LLMs for Deterministic Outputs

Interfaze released Structured Output Benchmark, scoring schema pass rate, types, and value accuracy across text, image, and audio. Each record has a JSON Schema and human plus LLM-checked ground truth; GLM-4.7 ranks No. 2 overall. The key bug is field-level value error: GPT-5.4 ranks 3rd on text and 9th on images.

Why it matters: Interfaze released the SOB benchmark, testing structured output across 5,000 text items, 209 images and 115 audio clips, checking whether field values are correct rather than just whether the JSON parses.