• Thesis on OpenAI's Fast-Follow Risk:
  • TypeSafe's Jev demonstrated a major breakthrough by offering rapid, single-pass probabilistic decisions instead of full generative text output.
  • Given Jev's explosive developer adoption on platforms like Vercel AI Gateway, the central argument is that frontier labs like OpenAI will quickly copy this architecture and provide their own version.
• Why Jev Is Vulnerable:
  • Architecturally, probabilistic zero-shot classification is not an insurmountable moat; it can be implemented by taking existing transformer backbones and tuning specialized output heads for calibrated scoring.
  • Frontier providers (like OpenAI) have massive data, compute advantages, and distribution channels to commoditize the format.
• The Integration Threat:
  • OpenAI can embed these fast decision and routing primitives natively inside their broader model stack and agentic pipelines.
  • Embedding internal snap judgments into frontier ecosystems would give developers a cohesive, single-provider solution where a standalone specialist model like Jev cannot easily follow or compete.
• Market & Product Takeaways:
  • While Jev's primary innovation is in packaging, ergonomics, and low-latency API design, defensibility against incumbents will hinge on speed of execution, ecosystem lock-in, and whether hyperscalers prioritize native first-party decision models.

AI Buzzwords EP.100故事线B把这条数据当作"harness决定Agent能不能被普通人用起来"这个判断迄今最有力的一手商业证据——内部98%说明模型能力本身没问题，外部17%/1%的巨大落差说明卡住普通用户的是产品/harness没做到位，不是模型不够聪明。同一篇报道里还有一个容易被忽略的细节：OpenAI的非工程团队一开始用Codex时，工具还在"敌视"他们——反复追问代码相关问题、提示"你这里有个空diff"，直到公司在2月到现在这段时间里把它做得更通用。这说明"让Agent普及"本身也是一个需要持续打磨的产品工程问题，不是模型发布后自动会发生的事。

EP.99 故事线B: OpenAI 将对齐证据的要求前移到「整个训练过程」，而不仅仅是训练后评估。这是方法论上的重大转变——从「训练后检查」到「训练中监控」，代表着 AI 安全实践的范式升级。

EP.99 故事线B: OpenAI 在这里承认主动放慢了训练速度——这是首次公开的「自我限速」声明。这不是被动合规，而是主动的安全判断。背后的含义是：Astra 或某个内部模型的网络能力评估触发了警报。

EP.99 故事线B: TAC（可信访问计划）本质上是一个「分级授权」实验——将高能力模型开放给被信任的安全研究员，而非完全限制。这正是 EP.99 预测的「分级授权成为默认」趋势的早期落地案例。

• Massive Net Losses: In 2025, OpenAI generated $13.07 billion in revenue but racked up $34 billion in total costs and expenses, resulting in an operating loss of $20.92 billion.
• One-Time Accounting Impact: Due to its transition from a non-profit to a for-profit entity, the company recorded a $41.55 billion loss from fair value changes in convertible interests and warrant liabilities. This brought the final net loss attributable to OpenAI to $38.53 billion.
• Year-over-Year Trajectory: Expenses and losses grew exponentially compared to 2024, when OpenAI brought in $3.7 billion in revenue against $12.48 billion in total costs, yielding a net loss of $5.09 billion.
• Core Expense Breakdown (2025):
  • Research and Development (R&D): $19.18 billion (up from $7.81 billion in 2024).
  • Cost of Revenue: $7.5 billion (up from $2.65 billion in 2024).
  • Sales and Marketing: $5.73 billion (up from $1.11 billion in 2024).
  • General and Administrative: $1.57 billion.
• Strategic Capital Flow & Microsoft Relationship: OpenAI paid Microsoft $17.2 billion in service fees during 2025 ($10.59 billion for R&D/model training and $6.047 billion for computing cost of revenue). By the end of 2025, OpenAI still had a remaining liability of $3.64 billion to Microsoft.
• Inbound Funding: Strategic partners provided substantial inflows; OpenAI received $867 million from SoftBank and $303 million from Microsoft in 2025.
• Remaining Cushion: As of the close of 2025, OpenAI held slightly over $50 billion in total assets, with nearly half of that cushion (~$25 billion) maintained as liquid cash reserves.

Hacker News Discussion

• R&D vs. Inference Costs: Commenters debate whether OpenAI can safely shift its massive R&D expenditure toward minimizing inference costs. While cheaper models like DeepSeek are heavily praised for personal and developer productivity, some argue stopping frontier model research means losing the structural race entirely.
• Diminishing Returns on Model Power: Users question whether a marginally smarter model justifies an exponentially higher cost. A central discussion point revolves around the financial viability of paying massive premiums for enterprise-tier models compared to utilizing low-cost API alternatives.
• The Math of Productivity Upgrades: A highly debated calculation suggests that even a 5% boost in productivity for a high-earning employee justifies hundreds of dollars in monthly subscriptions. However, critics counter that the financial surplus of that productivity is captured by companies and owners, rather than resulting in worker wage increases.
• The Path to Monetization: The consensus leans toward enterprise seat monetization (charging upwards of $2,000/month per corporate professional) and securing multi-billion dollar government contracts as the only viable business models. The inevitable integration of embedded or covert advertisements for free tiers is also viewed as highly likely.
• AGI as a Pseudo-Religious Goal: Several participants view Silicon Valley's relentless capitalization of unprofitable AI models as an irrational, faith-based pursuit of AGI (Artificial General Intelligence), comparing the narrative to religious prophecies.

这是一个重要的事实声明，需要核实OpenAI确实已提交IPO申请。文章提到这是'机密文件'，但没有提供更多细节，如申请的具体时间、预计上市时间或估值等信息。这些细节对于全面了解这一事件至关重要。

This is an underappreciated implication of AI company IPOs: going public doesn't just raise capital, it converts private decision-making into a domain subject to shareholder resolutions, proxy votes, and public disclosure requirements. The governance leverage that currently applies to Alphabet and Microsoft will extend to the frontier AI labs — a structural accountability shift that no amount of voluntary safety commitments currently provides.

引用OpenAI的案例提供了一个令人惊讶的事实：即使是AI领域的领导者也需要经历复杂而漫长的过程来构建有效的数据代理。这暗示了数据代理的成熟可能比市场预期的更晚，挑战了快速部署的乐观预期。

令人惊讶的是：即使是像OpenAI这样的AI领军企业，构建内部数据代理也是一个漫长而复杂的过程。这一事实揭示了当前AI技术在实际企业应用中面临的巨大挑战，挑战了人们对AI技术成熟度的过度乐观预期。

令人惊讶的是：OpenAI竟然给团队提供了极大的自主空间，让他们完全自由地探索AI编码的极限，这种开放的创新环境在大型科技公司中极为罕见，通常大型公司会严格控制研发方向。

OpenAI's ChatGPT will be added to the VLOSPs wrt search. Threshold is 45M amu's. OpenAI reported last fall 120M amu's. So this is a rather slow EC response I'd say.

OpenAI 的开源模型中存在一个专属的「版权拒绝机制」特征——这意味着版权合规行为是被明确编码进权重的，而非自然涌现的。更令人深思的是：同类竞争模型中不存在这个特征，暗示不同开发者对版权问题的训练决策存在根本性差异。

ChatGPT, next to adding ads, is now also inviting users to upload their contact list according to this web article. The purpose is unclear, is OpenAI building a social media platform? It is a EU wide GDPR violation though. Even if I were a user and opt-out others who don't might upload my personal contact details.

Adverts are coming to ChatGPT in the free-of-charge and lower paid tiers of subscription. Quelle surprise.

Where US healthcare fails, the amount of ChatGPT consultations grows. If anything expontentially is in play, it may well be exponentially worsening health outcomes for US population.

Author sees OpenAI losing their lead in 2025: Nano Banana Pro (Google) is a better image generating model Opus 4.5. better or equal than GPT5.2 Codex Max for coding Chinese labs have better open weight models Audio, Gemini Live API (google) is direct threat.

OpenAI mostly has better consumer visibility (yup, ChatGPT is the general term for LLMs, Aspirin style)

It is still strongest in consumer facing apps, but Gemini 3 is a challenger there.

international math olympiad style questions can be answered by OpenAI and Gemini models without tools nor having the challenges in their training data.

o for omni, as in multimodal outputs (text, image, sound?)

Wispr Flow uses both open (Llama) and closed LLMS, ao OpenAI . Server side though

They just completed a stricter regime by the DPA, and then immediately this comes up afterwards. Ouch. It surfaced bc they were monitoring network traffic volumes. That says to me they also know who did the uploading

Both these things, internal training and an AI operational code, are common in gov agencies. Here they are too late, and it's uncertain anyway they would have had effect. Any person who thinks nothing of uploading internal documents into a public website won't be held back by a rulebook they would not have read.

OpenAI requested to delete uploaded documents. Hardly possible I suspect, and definitely not something one can verify.

As result of the data breach, all OpenAI products have been banned internally. Only MS Copilot, embedded in their MS office suite is available.

GPT-OSS is by OpenAI. It is available locally in Ollama it seems in various versions.

100x seems a fantastic return. What are the upper reaches of unicorn investing in comparison to this? Many companies fail outright, so is it really a horrible investment?

This has all the hallmarks of a Ponzi scheme.

Hao, Karen. 2025. Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI. 1st ed. New York: Penguin Press. https://amzn.to/4o92MBs.

• NVIDIA’s latest earnings highlight a dramatic surge in AI demand, driven by a shift from simple one-shot inference to more complex, compute-intensive reasoning tasks.
• Reasoning models require hundreds to thousands of times more computational resources and tokens per task, significantly increasing GPU usage, especially for AI coding agents and advanced applications.
• Major hyperscalers like Microsoft, Google, and OpenAI are experiencing exponential growth in token generation, with Microsoft alone processing over 100 trillion tokens in Q1—a fivefold year-over-year increase.
• Hyperscalers are deploying nearly 1,000 NVL72 racks (72,000 Blackwell GPUs) per week, and NVIDIA-powered “AI factories” have doubled year-over-year to nearly 100, with the average GPU count per factory also doubling.
• To meet this unprecedented demand, more than $300 billion in capital expenditure is being invested this year in data centers (rebranded by NVIDIA as “AI factories”), signaling a new industrial revolution in AI infrastructure.

yeah, externalising the cost of getting it wrong at scale. Testing it in real world circumstances is extremely useful and needed, yet OpenAI's general public customers will mostly not show their own CT and assume any result is true (seen it happen a lot) and thus moving the cost externalisation further down the chain, where it is more likely to have negative real world consequences.

• openai use LiveKit to deliver realtime voice
• playground: https://cloud.livekit.io/projects/

While today’s LMs agents don't pose a serious risk, we should be on the lookout for improved autonomous capabilities as LMs get more capable and reliable.

~$50m = estimated training cost of GPT-4

OpenAI’s voice cloning AI model only needs a 15-second sample to work

As an ex-Viv (w/ Siri team) eng, let me help ease everyone's future trauma as well with the Fundamentals of Assisted Intelligence.<br><br>Make no mistake, OpenAI is building a new kind of computer, beyond just an LLM for a middleware / frontend. Key parts they'll need to pull it off:… https://t.co/uIbMChqRF9

— Rob Phillips 🤖🦾 (@iwasrobbed) October 29, 2023

[[Rolf Kleef]] [[Aldo de Moor]]

OpenAI call "democratic inputs for AI" 10 projects, 100k, 3 months

[[Rolf Aldo Common Ground AI consensus]] #2023/11/01

Minda Zetlin. "Bill Gates Says We're Witnessing a 'Stunning' New Technology Age. 5 Ways You Must Prepare Now". Inc.com, March 2023.

https://openai.com/blog/teaching-with-ai

https://openai.com/

Assistant is who answering questions?

User is who asking questions?

Default system message.

How to set up a good system message?

OpenAI API

A.I. Is Mastering Language. Should We Trust What It Says?<br /> by Steven Johnson, art by Nikita Iziev

Johnson does a good job of looking at the basic state of artificial intelligence and the history of large language models and specifically ChatGPT and asks some interesting ethical questions, but in a way which may not prompt any actual change.

When we write about technology and the benefits and wealth it might bring, do we do too much ethics washing to help paper over the problems to help bring the bad things too easily to pass?

What are the long term effects of fast breaking changes and gradual changes for evolved entities?

OpenAI has memorised MBA tests- when these are rephrased or certain details are changed, the system fails to answer

GPT4 knows the link to the coding exams that it was evaluated against but doesn't have "internet access" so it appears to have memorised this as well

OpenAI was only able to pass questions available before september 2021 and failed to answer new questions - strongly suggesting that it has simply memorised the answers as part of its training

https://openai.com/product/dall-e-2

DALL·E 2 is an AI system that can create realistic images and art from a description in natural language.

Whisper는 범용 음성 인식 모델입니다. 다양한 오디오의 대규모 데이터 세트를 학습하고 다국어 음성 인식, 음성 번역, 언어 식별을 수행할 수 있는 멀티태스킹 모델이기도 합니다.

텍스트 프롬프트를 기반으로 처음부터 이미지 만들기 새 텍스트 프롬프트를 기반으로 기존 이미지의 편집본 만들기 기존 이미지의 변형 만들기

https://platform.openai.com/

https://openai.com/blog/chatgpt

Educator considerations for ChatGPT<br /> https://platform.openai.com/docs/chatgpt-education

<small><cite class='h-cite via'>ᔥ <span class='p-author h-card'>Hypothesis</span> in Liquid Margins 38: The rise of ChatGPT and how to work with and around it : Hypothesis (<time class='dt-published'>02/09/2023 16:11:54</time>)</cite></small>

Origins of OpenAI

First a non-profit started with funding from Musk, Theil, and others. It has since transitioned to a "capped for-profit company".

Share-alike for restrictions, assuming that there will be proliferation of different choices of restriction sets.

I saw an AI Tutor demo as ASU+GSV in 2021 and it was still early stage. Today, the features highlighted here are yet to be manifested in powerful ways that are worth utilizing, however, I do believe the aspirations are likely to be realized, and in ways beyond what the product managers are even hyping. (For example, I suspect AI Tutor will one day be able to provide students feedback in the voice/tone of their specific instructor.)

Johnson, Khari. ‘AI Could Soon Write Code Based on Ordinary Language’. Wired. Accessed 21 June 2021. https://www.wired.com/story/ai-write-code-ordinary-language.

OpenAI和其他研究人员已经发布了一些能够识别AI生成的文本的工具。它们使用类似的人工智能算法来发现文本中的蛛丝马迹。目前还不清楚是否有人利用这些来保护在线评论平台。Facebook拒绝透露是否在使用这类工具；谷歌和Twitter没有回应置评请求。

去年6月，OpenAI 发布了一个更强大的文本生成程序，称为 GPT-3。到目前为止，它只向少数人工智能研究人员和公司开放，一些人开发了有用的应用程序，比如从要点生成电子邮件信息的程序。当GPT-3发布时，OpenAI在一份研究报告中表示，尽管它已经意识到Weiss的研究，但没有看到GPT-2被恶意使用的迹象。