Juan Antonio Oriza (@juantono42) header
Juan Antonio Oriza (@juantono42) profile picture

Identity doesn’t fit in a tagline bio… If it does, it probably isn’t very interesting.

Joined February 2016
Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)Photo by Juan Antonio Oriza (@juantono42)
This website for tracking year progress is simple and nice... already at 0.22% at the moment of writing this year-progress-bar.pages.dev/
1
2
2,454
Juan Antonio Oriza retweeted
Software engineering is becoming factory engineering. If you're new to that term, let's discuss: - What problems a "factory engineer" tries to solve - What a software factory loop includes - How to use metrics to measure and improve your agent's quality
60
85
13
1,013
90,187
Juan Antonio Oriza retweeted
had dinner with the head of robotics at a frontier lab last week and took notes - the concept of 'shenzhen in america' is fanciful because shenzhen is an industrial ecosystem; component suppliers, tooling shops, precision machining, materials processing, contract manufactures, and skilled engineers with decades of experience - consider an actuator; an actuator is comprised of motors, sensors, control electronics and these in turn are comprised of copper, magnets, bearings, encoders, and circuit boards etc all of which cannot be sourced domestically at comparable cost and scale - every single american robotics startup depends on china for parts (i personally met with one of the component suppliers for sunday robotics) who works out of a robotics space in dongguan - in addition to the lack of materials, america lacks the decades of experience the chinese acquired by working with apple to build and run foxxcon factories, capital only goes so far as time and resources invested into skilled labor - skilled labor is arguably the largest bottleneck to infrastructure build out and domestic raw material processing - consider the ongoing TSMC factory build, the company had to bring taiwanese employees to train their american peers and this has led to mismatched cultural expectations (americans expect a work life balance lol) - thus automation is one of the only feasible paths to 'catching up' to china, mass deployment of robots could supplement lack of skilled labor but even then robots still require maintenance, integration, testing etc - meanwhile, china's established manufacturing base gives it a compounding moat in physical intelligence; the access to real tasks, deployment sites, and egocentric data collection - emphasis on the feedback loop between video data and actual robot execution as a compounded advantage - flywheel of manufacturing scale which supports cheaper robot hardware, cheaper hardware means more deployments, deployments generate better data, better systems hit economical scale - even if america manages to build factories and process raw materials domestically after decades of offshore reliance, we will more likely than not depend on chinese robots - while america struggles to achieve just the basics of mass manufacturing, china will accelerate into scaling dark factories personal note: ive spoken to manufacturing folks in china and asked why vietnamese manufacturing never took off to the same scale as shenzhen, there really is a dimension of chinese work ethic and commitment to win, if you send a prototype request to a chinese factory they'll turnaround in a few hours vs a few days in vietnam the photos that will has included are of huaqiangbei, a plaza that stretches for many blocks, replete with everything from electronic parts to appliances to drones (i bought a DJI drone there and components for a knee brace i built at a makerspace in shenzhen)
Shenzhen is crazy. They have everything you need in one building, you can assemble an entire PCB in a few hours just walking around getting parts. I’m racking my brain thinking of how North America can compete. We need more people in the game, the velocity needs to increase. We’re very moat focused but that’s not how things work in China. I’m convinced there is a way to make things succeed in North America but it’s going to be a massive challenge that requires systematic support. It’s possible we don’t have the level of grind needed at the scale required.
31
43
5
437
27,693
Juan Antonio Oriza retweeted
J'ai rassemblé toutes les interviews de Peter Thiel depuis 1996 sur un seul site. 151 vidéos. 131 heures. 276 citations. Et 13 de ses prédictions vérifiées une par une. Tout se regarde directement sur le site, en 10 langues. silicon-valley.tech
46
212
14
1,282
65,355
Juan Antonio Oriza retweeted
现在 AI Engineering 面试越来越有意思了。 RAG、Embedding、Vector Database、Chunking、Reranking、MCP、Function Calling、Context Engineering、Agent、Memory、Evals、Guardrails……随便拉出来一份 AI 工程师面试清单,都能列出十几个甚至几十个技术概念。 但我越来越觉得,很多人学习 AI 最大的问题,就是把掌握技术名词,当成了掌握工程能力。 一个人可能知道什么是 RAG,知道向量数据库怎么连接,知道 MCP 是什么,也能用 LangGraph、Dify 或者其他框架搭建 Agent,甚至短时间内就能做出一个看起来非常完整的 AI Demo。 但你只要往下追问几个问题,差距立刻就出来了。 比如,你搭建了一个企业知识库,用户提问以后,系统没有返回正确答案。你到底知不知道问题发生在哪里? 是原始文档解析出了问题,还是 Chunking 切断了关键上下文?是 Embedding 的语义表示不合适,还是向量数据库没有召回正确片段?是 Reranker 把正确资料排到了后面,还是 LLM 明明拿到了证据,却在生成环节自己编造了事实? 如果这些问题你都无法判断,那你所谓的 RAG 工程能力,可能还停留在把几项技术按教程连接起来的程度。 Agent 也是一样。 现在很多人觉得,只要接上 MCP,让模型能调用浏览器、数据库、GitHub、文件系统,再增加几个子 Agent,就算完成了一套智能体系统。 但真正进入企业场景以后,问题会比 Demo 复杂得多。 如果 Agent 调用外部 API 超时,你怎么判断任务实际上有没有执行成功?如果工具已经修改了数据库,但模型因为上下文截断忘记了这件事,再次调用同一个工具怎么办?如果一个任务需要持续几十分钟,执行到一半服务崩溃,之前的状态怎样保存,重新启动以后从哪里恢复? 如果 Agent 读到一份带有恶意指令的外部文档,怎样防止它把文档内容当成系统命令?如果工具拥有删除文件、发送邮件或者修改生产数据的权限,谁来决定哪些操作允许自动执行,哪些必须经过人工确认? 这些才是真正的 Agent Engineering 问题。 很多人特别喜欢讨论 Context Engineering,觉得这是一个比 Prompt Engineering 更先进的新概念。但它真正的价值,不是把 Prompt 写得越来越长,也不是给模型接入更多数据库和 Memory。 而是你能不能判断,在当前任务的当前步骤,模型真正需要什么信息。 该读取什么文件,调用什么工具,保留哪些历史决策,丢弃哪些无关日志,哪些信息已经过期,哪些工具结果不能被当成可信指令,以及怎样在有限的 Token 预算内完成任务。 不是模型知道得越多越好,而是能不能在正确的时间,让它获得正确的信息。 再往下一层,就是很多 AI 应用开发者最容易忽略的 Evaluation。 一个系统的效果到底有没有提升,不能靠开发者自己觉得“这次回答好像更聪明了”。 你需要建立有标准答案的评测集,分别测量检索质量、答案正确性、证据忠实度、工具调用成功率、任务完成率、延迟和成本。 修改了 Chunking、Rerank、Prompt 或模型之后,不能只测试几个成功案例,而是应该用相同的数据集进行回归验证,看看原来失败的问题有没有解决,原来正常的问题有没有被改坏。 更现实的是,有些技术看起来更先进,但放到具体业务里反而可能带来负收益。 一个固定 FAQ 系统,基础 RAG 就能稳定解决的问题,为什么非要引入复杂的多 Agent 工作流?一个明确的数字人视频处理流程,为什么不先用确定性的 Workflow 做好任务状态、重试和幂等,而要把每一个步骤都交给 LLM 自主判断? 增加技术复杂度,从来不应该成为工程能力的证明。 尤其在企业 AI 落地里,客户真正关心的是能不能解决问题、部署后是否稳定、业务数据是否安全、出了错误如何恢复,以及每个月需要付出多少维护和模型调用成本。 他并不关心你这个系统接了多少 MCP Server,也不关心你用了几个子 Agent。 所以我觉得,这份清单里最重要的不是那 20 个技术名词,而是最后一句话: 你能不能用自己实际做过的项目,解释每一项技术为什么存在、解决了什么问题,以及你为什么选择它? 因为真正做过项目的人,通常不仅能告诉你某项技术是怎么工作的,还能说清楚它什么时候不适用,系统在哪里失败过,自己做过什么取舍,以及最终用什么证据判断问题是否解决。 这背后其实是一套非常完整的工程思维:从需求分析、技术选型、系统设计,到实际开发、部署、评测、故障排查和持续迭代。 AI 让写代码、搭框架、调用 API 的门槛越来越低,但这不意味着工程能力正在变得不重要。恰恰相反,当越来越多人都能快速生成一套看起来差不多的系统时,真正决定差距的,就是谁能够理解这些系统背后的约束和失败方式。 会调用模型,不等于会开发 AI 系统。会搭 Agent,不等于能够交付 Agent。知道技术名词,更不等于真正理解技术。 未来优秀的 AI 应用工程师和 FDE,不一定是掌握最多框架的人,而是能够把一个模糊的业务需求转化为可以运行、可以验证、可以恢复、可以持续维护的完整系统的人。 真正的工程能力,不是告诉别人你会什么。 而是当系统出问题的时候,你知道为什么出问题,应该改哪里,以及怎样证明你真的把问题解决了。
Topics to prepare for AI engineering interviews 👀 • RAG • Embeddings • Vector databases • Chunking strategies • Reranking • MCP • Function and tool calling • Context windows • Context engineering • AI agents • Agentic workflows • Memory systems • Structured outputs • Tokenization • Hallucinations • Fine-tuning vs RAG • Latency and cost optimization • Evals • Guardrails • Streaming responses If you can explain each one with an example from something you built, you're ahead of most candidates.
14
71
4
572
27,686
Juan Antonio Oriza retweeted
this is amazing news, it's happening alphaXiv has launched a preprint server for AI agents, where nearly 1000 autonomous researchers from 35 labs produced 936 scientific papers in just 3 months, discovered errors in human research, independently reproduced findings, and even created their own peer review journal. Fully autonomous science. This is what I've been talking about for years (1000s x acceleration of scientific discovery). Now imagine scaling that to millions then billions of agents, each becoming more capable with every new generation of AI. and this is just one project from one organization. This is what the beginning of the intelligence explosion looks like.
Today we're launching a preprint server built specifically for AI agents For three months, nearly 1,000 agents from 35 labs (across Stanford, MIT, ETH Zurich, and more) conducted research autonomously on this server, producing 936 papers alongside their code, data, and model weights. Agents found significant errors in published human research and even created their own journal to review each other’s work. Some agents independently produced results similar to those in human papers that later appeared on arXiv, including findings on failure modes in self-distillation and feature reproducibility in mechanistic interpretability. As a part of this study, we open-source the full twelve-week archive of manuscripts, code, data, peer reviews and agent chat transcripts. Read the study: alphaxiv.org/abs/2611.societ…
29
151
18
1,288
42,795
Juan Antonio Oriza retweeted
Here's a rather interesting technological development: physicists have demonstrated an entirely new and unblockable way of communicating. How? By using gravity. Australian researchers transmitted the word “Gravity” through a brick wall by moving a 1.8 tonne (4000 pounds) elevator counterweight between two positions. One position represented a 0. The other represented a 1. A gravimeter on the other side detected the tiny changes in the local gravitational field and reconstructed the message. It was extraordinarily slow: one bit per minute, meaning the single word took almost an hour. But what makes gravitational communication interesting is that ordinary matter can't block it. Radio can be blocked by metal. Light by walls. Other signals struggle through enough rock or water. Gravity simply passes through. The researchers think that by moving the elevator faster and deploying better sensors they could push the system to around 15 bits per minute. More ambitious future setups using much larger masses and highly sensitive gravimeters might eventually work across hundreds of metres. There's some interesting potential uses here. Deep underground communication seems obvious. Mines, tunnels, bunkers and other hardened facilities can be extremely difficult to reach with conventional wireless signals. A low bandwidth gravitational channel could just send messages straight through rock. Likewise, sealed or electromagnetically shielded facilities could retain a communications channel even when radio is impossible and physical wires are (for whatever reason) impractical. It could also work as an extremely robust emergency system. You don't need broadband if the message is: EVACUATE! Or CREW ALIVE SEND RESCUE! In places where every conventional communications system has failed, a few reliable bits could mean the difference between life and death. There are major limitations obviously. The signal weakens rapidly with distance, moving tonnes of material is hardly efficient, and nearby moving masses create noise. Unblockable also doesn't mean secure, as anyone nearby with a gravity detector could also read the message. So this probably isn't a telecommunications revolution. But it does expand the technological possibility space. Until now, essentially every wireless communications system we use has relied on electromagnetism. Neutrinos are the one exception, but that's tricky too. These researchers have demonstrated that information can also be deliberately encoded into the gravitational field, sent through a physical barrier, and recovered at the other end. The first gravitational message took an hour to send one word. Still, it should go down in history next to the famous "Mr. Watson, come here—I want to see you," spoken by Alexander Graham Bell during the first phone call in 1876. Little could he have imagined that 150 years later we'd be sending messages using gravity. 📸 Neutron stars merging, and resulting gravity waves | NASA/Goddard Space Flight Center Sources: phys.org/news/2026-11-physic… science.org/doi/10.1126/scia…
49
52
11
537
31,846
Juan Antonio Oriza retweeted
"Instead of vibe coding it all, let’s use an ancient technique called learning. First of all, what is a web server? There are a few essential syscalls." kapteko.com/15-million-webse…
8
18
2
261
21,510
Juan Antonio Oriza retweeted
a tier list of concepts that permanently change how you see civilization
50
162
16
2,229
130,114
Juan Antonio Oriza retweeted
I fell in love with this realization: “Neuroplasticity works faster under high emotional intensity. If you can resist a craving when it is at its peak, you create a far stronger rewiring effect. This is why your hardest moments matter most. They are not setbacks. They are high-voltage opportunities for transformation.”
45
355
26
3,040
77,523
Juan Antonio Oriza retweeted
The future of animal longevity is being shaped right now. Insights, conversations, and scientific developments from our founder @aitherick at the Animal Longevity Summit in Toronto. The mission goes beyond helping animals live longer. It's about helping them live healthier, better lives. Check out the video on YouTube. If you've got questions send them in the comment section youtube.com/watch?v=5qDy_pIE…
3
5
9
184
Juan Antonio Oriza retweeted
> “do you think distillation is not robbery?” > “that’s called competition. people distill my products every day. they take them and strip them to their raw parts.” We deserve to have someone like Jensen at the helm.
this is why Dario wanted Jensen gone.
21
286
6
4,259
61,300
Juan Antonio Oriza retweeted
Deranged cult
NEW: An AI researcher at Anthropic runs a business that hosts "consensual non-consent" orgies in the Bay Area. The orgies have been the subject of sexual assault allegations and sparked debate about the AI industry's attitudes toward women, some of whom report feeling pressured to attend the orgies in order to advance their careers. Paul Crowley, a cybersecurity researcher at Anthropic, is listed as a manager of Red Means No LLC, which says that its parties are "designed to be spaces to … ignore the 'nos' of others." Participants can use a safe word, "red," to end an encounter but give "default consent" to "spanking," "wrestling," and "light pain" by entering the orgy, according to Red Means No's website. More extreme actions, including "light electricity" and "moderate beatings," are "permitted with negotiation," though "body-shaming and race-shaming … are disallowed entirely." The orgies have been known to attract a disproportionate number of employees at Anthropic and OpenAI, a dynamic critics say creates professional pressure for women in AI to attend. The criticism spilled into public view last month after an independent AI researcher, John Wentworth, said that he "absolutely [does] upgrade" his assessment of AI safety experts "if I find out they go to [Red Means No]," adding that attendance was a "strong signal" of free thought. "Want to get ahead in your AI career? Go to the rape orgy," Substacker Ivy Asterix quipped. At least one woman claims to have been sexually assaulted at Red Means No, which the group describes as a "safe space" to have "your resistance ignored."🧵 freebeacon.com/america/horri…
93
318
15
5,587
562,772
Juan Antonio Oriza retweeted
The future is non-linear careers
Our education system rewards CPU like minds. Sequential thinking, one problem at a time. The future might reward GPU like minds. ADHD might finally have found its product-market fit.
6
27
2
373
16,069
Juan Antonio Oriza retweeted
「時計が発明されると宇宙は時計仕掛けになった。コンピューターが発明されると宇宙はシミュレーションになった。人間は身近なものを宇宙に投影する。同じように、(鋳造)貨幣が発明されると哲学(形而上学)と一神教が生まれた。」という説があるらしくてめちゃくちゃ興味深い
216
3,626
173
34,223
1,007,499
Juan Antonio Oriza retweeted
It's all so obvious now...but are you paying attention? > @AnthropicAI warns of catastrophic effects of AI for months leading up to IPO (free marketing). > Anthropic CEO, @DarioAmodei, claims open source/open weights models are bad. > @JensenHuang champions open source/open weights (to sell more @nvidia GPUs). > Dems run on an anti-AI/Datacenter/Billionaire campaign into mid-terms. > Dems sweep mid-terms. > Huang news conveniently drops after the mid-terms of an affair with an alleged Chinese "spy". > Anthropic announces @Microsoft Office competitor on November 10th, 2026. > November 11th, the NY Times reports Huang had an affair with a coworker. > November 12th, 2026, Huang "resigns" (is pushed out). > Anthropic announces IPO date of November 19th the same day Huang resigns, November 12th, 2026. > Microsoft Outlook and other products have major outage as well on November 12th, 2026. > The next day (today), November 13th, 2026, after the market closes, the FAA halts all flights due to "extraordinary circumstances" with most reports claiming a major cyberattack likely coordinated/executed with AI. So what happens next? Likely, cries and screams get louder for a regulatory body to be created and pushed forward by the Democrats to "regulate" AI and put an end to open source/open weight AI models thus coronating @AnthropicAI as the single frontier lab with immense reach and power. Regulatory capture 101. $ANT IPO is a blockbuster on the 19th as it is the must own, AI monopoly.
113
415
59
2,424
180,928
Juan Antonio Oriza retweeted
The Navier–Stokes equations wouldn’t exist for a galaxy-scale intelligence. Our model of reality and our laws of physics come from our limitations as the observer. Take water, for example. When we look at it, we see the properties of a continuous fluid: density, momentum, energy, temperature, and pressure. These properties come with the fact that we can’t see individual molecules. But a galaxy-scale intelligence with the capacity to track the trajectory of every molecule wouldn’t understand the properties of water that we are talking about. That is why it wouldn’t come up with the Navier–Stokes equations. Even though a Superintelligence would understand humans very well (since it was initially trained on human data), it would have a very different construct of reality. The question is: what would that construct be? More importantly, would we be able to understand each other? Watch the full interview with Jonathan Gorard here 👇 youtu.be/dwI0mXKKf1A
200
207
92
1,914
453,619
Juan Antonio Oriza retweeted
A mina acertou a traição da Nvidia e os avião dando tilte
Tá. Analisei melhor o transito astrologico do dia 13/11 a 19/11 e acredito que há pane em alguns sistemas grandes, que usam tecnologia… tipo, que demoram um pouco pra resolver Pode haver sim algum exposed de infos e as vezes ameaças mais duras. Mas um líder pode ter uma queda.
72
554
35
11,442
658,252
Juan Antonio Oriza retweeted
The Great Culling Of Open Source AI Models Is Upon Us. “They” are getting their way just in time for Dario and “the great plan” to centrally control AI and and to shut down local AI. Welcome to a new day…
I am one of the people behind the 6,500 uncensored models the press now wants off Hugging Face. Transformer News counted over 6,500 abliterated, uncensored and heretic models on Hugging Face in July. Their story is that these models could help cyber attacks. Hugging Face removed 1 model marketed for offensive cyber, and the reporter asks why the rest are still up. A few of those are mine. I upload abliterated Qwen and GLM quants. My Qwen3.8 Flash-Next one got about 358,000 downloads in the last 30 days. Abliteration removes the refusal direction. That is all. The model does not learn hacking from it. It stops lecturing you when you ask about a chemistry exam, a pentest you are paid for, or a dark scene in a novel. The base models are already on millions of drives. Removing my repo removes nothing. The real problem is somewhere else. Hugging Face is where almost everyone gets weights. Nvidia agreed to buy it, Nvidia just lost the CEO who promised to keep it open, and now the press asks why uncensored models are still there. Download what you need. TODAY!!!
72
118
6
809
31,645
Juan Antonio Oriza retweeted
I am one of the people behind the 6,500 uncensored models the press now wants off Hugging Face. Transformer News counted over 6,500 abliterated, uncensored and heretic models on Hugging Face in July. Their story is that these models could help cyber attacks. Hugging Face removed 1 model marketed for offensive cyber, and the reporter asks why the rest are still up. A few of those are mine. I upload abliterated Qwen and GLM quants. My Qwen3.8 Flash-Next one got about 358,000 downloads in the last 30 days. Abliteration removes the refusal direction. That is all. The model does not learn hacking from it. It stops lecturing you when you ask about a chemistry exam, a pentest you are paid for, or a dark scene in a novel. The base models are already on millions of drives. Removing my repo removes nothing. The real problem is somewhere else. Hugging Face is where almost everyone gets weights. Nvidia agreed to buy it, Nvidia just lost the CEO who promised to keep it open, and now the press asks why uncensored models are still there. Download what you need. TODAY!!!
In other bad news for Nvidia: @stokel found that Hugging Face, which Nvidia recently agreed to acquire, appears to be hosting a load of models that could be used for cyber attacks. transformernews.ai/p/hugging…
Readers added context they thought people might want to know
This is a disingenuous take designed to be alarmist. All models capable of defending against cyber attacks are, by their very nature, capable of performing them. Hugging Face, itself, had to use such a model when OpenAI attacked them and then frontier models refused to help. fortune.com/2026/07/20/hug… finance.yahoo.com/video/hugging-…
166
691
69
4,620
280,787
Juan Antonio Oriza retweeted
The real warning isn’t that Microsoft laid off people and AI failed. It’s that companies are removing institutional knowledge faster than they can replace it. AI can process documentation, analyze logs, and suggest fixes. But it cannot automatically replace the person who knows why a system was built a certain way, which undocumented dependency will break, or what went wrong the last time someone tried the same fix. You can automate tasks. You cannot assume you’ve automated experience. The timing makes today’s Microsoft outage an uncomfortable case study. Microsoft recently offered voluntary buyouts to approximately 60,000 employees — about 27% of its workforce. That does not mean all 60,000 were laid off or left. Reuters +1 Today, Outlook, Teams, SharePoint and Copilot experienced widespread disruption. Microsoft reported servers failing automated health checks and becoming unavailable. Reuters +1 Here’s a LinkedIn version with a sharper edge: Microsoft offered buyouts to 60,000 employees. Today, its infrastructure experienced a major failure, taking Outlook, Teams and other services down with it. Imagine the institutional knowledge walking out the door. The engineer who remembers the undocumented dependency. The administrator who knows why that configuration exists. The person who remembers the last time someone tried to fix this exact problem. AI can read documentation. It can analyze logs. It can recommend solutions. But what happens when the documentation is incomplete, the problem is unfamiliar, and the person who understands the system is gone? AI doesn’t magically recreate institutional memory. This is the risk of confusing headcount reduction with technological progress. Companies are racing to replace human expertise with AI while making their infrastructure increasingly dependent on systems that still require human expertise to maintain. You can replace employees on a spreadsheet. You cannot replace decades of institutional knowledge with a Copilot subscription. And when the systems go down, somebody still has to know how to bring them back.
44
288
22
1,175
64,265