热门科技推文速览——2026年8月23日

今日科技动态中，人工智能基础设施与智能体工具成为关注焦点。据报道，Anthropic正探索以2万亿美元估值进行一场重磅首次公开募股（IPO），英伟达Vera Rubin系统也已开始出货。开发者重点关注了启动速度提升25倍的Codex CLI、专用Grok机器人、价格亲民的AI视频生成工具，以及表现出色但来历神秘的Ox Alpha编程模型。与此同时，数据中心正成为强劲的增长引擎，其中包括建设太空数据中心的计划；Linux桌面系统雄心再起，加之开源ThreeUI库的推出，也为工程技术社区注入了新的活力。

1. wallstengine (Group Score: 190.0 | Individual: 35.6)

Cluster: 7 tweets | Engagement: 336 (Avg: 194) | Type: Tech

ANTHROPIC COULD TARGET $100B+ IPO AT$ 2T VALUATION

Anthropic’s bankers have discussed an IPO that could raise more than $100B and value the AI company at roughly$ 2T, per NYT.

If completed at those levels, it would be the largest IPO ever, topping SpaceX’s $85.7B raise at a$ 1.77T valuation.

Anthropic was valued at $900B earlier this year and recently surpassed$ 65B in annualized revenue, up from $9B at the end of 2025.

The company is expected to release its IPO prospectus in the coming weeks, with a potential listing in the coming months.

See 6 related tweets

@IPONewsroom_: ANTHROPIC COULD SEEK RECORD $100B+ IPO

Anthropic’s bankers have told potential investors the AI sta...

@WOLF_Financial: Anthropic could seek to raise $100B + at a$ 2 Trillion valuation in its upcoming IPO, per NYT.

Its ...

@chandrarsrikant: RT @srimuppidi: 🍦SCOOP: Anthropic could raise more than $100B in its upcoming IPO at a$ 2T valuation...
@negligible_cap: *ANTHROPIC COULD AIM TO RAISE $100B IN IPO: NYT

“In recent discussions with potential investors, An...

@Techmeme: Sources: Anthropic's bankers said the company could raise $100B+ in its IPO, which could value it at...

2. Hesamation (Group Score: 173.5 | Individual: 35.3)

Cluster: 8 tweets | Engagement: 96 (Avg: 246) | Type: Tech

VERA RUBINS ARE SHIPPING.

Nvidia said back in January that Microsoft will be among the first for Rubin deployment (along with AWS, Google, CoreWeave). CoreWeave got one running in June, Google got racks by July too.

looks like Rubins are starting to roll out at scale. https://t.co/1hfqP7QcOo\n\nQT @satyanadella: Delivery day at our Microsoft DCs as the first production Vera Rubins arrive. A huge thank you to our partners at @nvidia and our Azure hardware and datacenter teams for all the incredible work that brought us to this milestone! https://t.co/pBqKArGO1N

See 7 related tweets

@benitoz: Vera Nice!

No delay

I can enjoy the weekend.\n\nQT @satyanadella: Delivery day at our Microsoft DC...

@zephyr_z9: OpenAI & Microsoft have started to get the initial VR200 batches now. MP starts soon\n\nQT @saty...
@ShanuMathew93: Vera Rubins out in the wild now 👀\n\nQT @satyanadella: Delivery day at our Microsoft DCs as the firs...
@tunguz: One Vera Rubin could fix me.\n\nQT @satyanadella: Delivery day at our Microsoft DCs as the first pro...
@DivesTech: RT @satyanadella: Delivery day at our Microsoft DCs as the first production Vera Rubins arrive. A hu...

3. NeuraFlowAix (Group Score: 160.0 | Individual: 34.0)

Cluster: 7 tweets | Engagement: 114 (Avg: 44) | Type: Tech

Ready to create without boundaries? CapCut provides the easiest editing tools combined with the best AI. The kicker? Seedance 2.0 Fast is now available for as low as $0.01 per second.

💻 Web: https://t.co/WbObOfftHV 📱 App: https://t.co/Ri6f3fQxo1

#CapCut #Seedance25 #CapCutai #CapCutDidThat

See 6 related tweets

@TheoBuildsAI: When the tool is simple, you can focus on the creative idea. 🎥

CapCut makes AI video creation easy ...

@ethancole_ai: Only CapCut can bring you this level of quality this easily. Seedance 2.0 Fast is now available for ...
@alexaiworks: A new era of video creation! CapCut is proud to offer Seedance 2.0 Fast for as low as $0.01/sec. We ...
@NovaIAHQ: We didn't just lower the price; we raised the bar. Seedance 2.0 Fast is now available for as low as ...
@zaynmcps: The secret weapon of top creators? The perfect workflow. CapCut offers the simplest editing tools ba...

4. heyrobinai (Group Score: 116.5 | Individual: 33.2)

Cluster: 4 tweets | Engagement: 43 (Avg: 133) | Type: Tech

How to get dangerously good at Grok Bot:

“personal assistant” is not a job. give each bot one recurring responsibility, clear inputs, and one result it owns.

brief every task with five things → outcome, sources, constraints, deliverable, and the point where the bot must stop for review.

start with an attached document and no login. test whether the bot stays inside scope, cites its evidence, respects the stop condition, and returns something you can use.

run the task manually three times. every run should expose one failure that becomes a lasting instruction.

when the first report is overwhelming, do not restart. tell the bot exactly what to keep, remove, rank, shorten, or reformat.

once the workflow works, save it as a skill. test the skill on another safe example, then schedule the fourth run.

deliver the result somewhere you already check. an automated report inside an app you never open is still useless.

start with read-only access. require approval before the bot sends messages, spends money, cancels anything, deletes data, or touches production.

for business bots, use this loop → signal → gather context → rank the opportunity → draft the action → human approval → execute → learn from the result.

pick the pain already costing you the most time, money, or attention. copy the closest build, run it three times, then schedule the fourth.

i broke down 21 bots using this exact system. every build includes its trigger, sources, deliverable, approval gate, and a starter prompt you can steall ↓\n\nQT @heyrobinai: https://t.co/q6DZU0MP9k

See 3 related tweets

@rewind02: grok bot is turning into the fastest UGC pipeline nobody's talking about yet...

not because it gene...

@milesdeutscher: The ULTIMATE guide to Grok Bot.

In 15 minutes, I'll teach you how to master the most powerful agent...

@elonmusk: It’s that easy to use Grok @Bot\n\nQT @Lummox_eth: I launched Grok Bot for 24 hours, gave it 3 jobs,...

5. PaulSolt (Group Score: 115.0 | Individual: 56.9)

Cluster: 3 tweets | Engagement: 2538 (Avg: 70) | Type: Tech

RT @dhh: THE ANNOUNCEMENT: We’re going to make the prophecy of The Year of Linux on the Desktop come true. All the pieces are now in place. Time to go all in! https://t.co/zuf5h9ClWX

See 2 related tweets

@AskYoshik: Imagine being so successful and rich that your side quest is now building your own Linux distro beca...
@nummanali: This is wonderful in so many ways

If you were to become free of work, well connected, and rich - w...

6. trevin (Group Score: 103.1 | Individual: 39.1)

Cluster: 3 tweets | Engagement: 284 (Avg: 56) | Type: Tech

RT @charliermarsh: In the latest Codex CLI release, I redid the lifecycle to make codex startup instant.

It's now ~25x faster and immediately responsive. https://t.co/F01ohFN1T1

See 2 related tweets

@mitsuhiko: I like this trend. These labs have some of the best engineers with agency just doing good work!\n\nQ...
@josevalim: Can confirm, visibly better on day to day usage!\n\nQT @charliermarsh: In the latest Codex CLI relea...

7. KyleReidhead (Group Score: 102.3 | Individual: 29.6)

Cluster: 4 tweets | Engagement: 36 (Avg: 27) | Type: Tech

RT @KyleReidhead: The USA hates data centres, meanwhile they're becoming one of the best growth drivers in the country!

The mainstream media needs to report on THIS rather than the fake news they have been spreading (please share this!)

The data is CLEAR, counties with an OPERATING data centre crush the US avg. on home value, employment and job growth:

New housing units built: +50% vs +1% Home values: +16% vs +7% Unemployment: 3.4% vs 4.2% Job growth: +3.8% vs +1%

And the tax side is fantastic too:

In Loudoun County, Virginia, data centers sit on about 4% of commercial parcels and throw off 38% of the county's entire general fund revenue. More than $100 million of NEW revenue/year!

Across northern Virginia data centers paid roughly $1.3 billion in local property taxes in 2024 alone (and this was as the boom was just getting started!)

The other states problems have no weight to them too. They use too much water?

US golf courses use roughly 2.08 billion gallons a day against data centers at about 449 million

Electricy costs are rising from Data Centres? Also not true. The opposite has been happening as Data Centers help to improve power generation within states or supply their own and sell the extra back to the grid.

And the last point to make here: Data Centers are what power AI.

The technology that is helping to treat cancer (as we saw with the recent Moderna announcement)

The technology that's increasing earnings across many companies in the USA and allowing them to hire more people

The technology that is allow anyone build a company cheaper and faster than ever before. AI is a net positive to the world and so are the data centres we need to power it

Don't let mainstream media continue to share bullshit fake news. Help share this content and spread the truth!

Don't forget to give me a follow @kylereidhead for more insights on AI and markets

See 3 related tweets

@ChristopherHale: .@a16z and others in Silicon Valley — this is a much better way to engage in this conversation with ...
@LoganDobson: Really appreciated the chance to go on @MTSlive because it helped me articulate something I’ve been ...
@TaylorLorenz: RT @LoganDobson: Really appreciated the chance to go on @MTSlive because it helped me articulate som...

8. mikenevermiss (Group Score: 98.6 | Individual: 39.3)

Cluster: 3 tweets | Engagement: 107 (Avg: 52) | Type: Tech

this is crazy.

someone built a full AI research team with Grok Bot for about $200/month.

a traditional Wall Street research setup can cost around $294,000/year:

« $27,000 - Bloomberg Terminal$ 22,000 - Refinitiv $50,000 - sell-side research$ 15,000 - AlphaSense $180,000 - junior analyst»

now, 6 AI agents can handle the same kind of research for about $2,400/year.

each agent has one simple job:

FILINGS - reads company reports and finds anything unusual.
EARNINGS - reads earnings calls and spots important changes.
SECTOR - watches different industries for important news.
INSIDER - tracks insider buying and big fund positions.
CHATTER - watches X for unusual activity around a company.
CHIEF OF STAFF - reads all the reports and sends you the most important findings.

they work overnight and send the final report before the market opens.

so instead of spending hours reading everything yourself, you wake up with the important information already organized.

and the best part is...

the person who built it published the prompts and the full setup.

the video below shows you how to build the whole thing yourself.\n\nQT @mikenevermiss: this guy just released a detailed 25-minute guide on Grok Bot, and this might be the best free course i’ve seen so far.

Grok Bot gives you an army of AI agents that can work for you 24/7.

set it up properly and you unlock some serious superpowers.

the video covers everything from setup and use cases to plugins and what makes Grok Bot so powerful.

bookmark this one.

See 2 related tweets

@poteto: RT @gregisenberg: Grok Bot might be the first tool that lets one non-technical person run an entire ...
@RoundtableSpace: THIS WALL STREET RESEARCHER REPLACED A $294K SYSTEM WITH 6 GROK BOTS THAT HUNT FOR ALPHA 24/7 AND SA...

9. omarsar0 (Group Score: 94.8 | Individual: 33.8)

Cluster: 3 tweets | Engagement: 212 (Avg: 49) | Type: Tech

Recommended agent skill.

/eli5 works great for visualizing deep technical concepts!

I find it speeds up my thought process and helps me collaborate better with agents.

Try /eli5 with Pi and Ox Alpha in our harness playground: https://t.co/7ivyJJ9xu2 https://t.co/QCA1J9PBoG\n\nQT @trq212: a skill people at Anthropic have been using a lot recently: ELI5

/eli5 <what you want explained>

"explain like I'm someone who knows nothing about this topic, using a HTML artifact with big pictures and few words" https://t.co/OZqzjAyFdT

See 2 related tweets

@robinebers: RT @trq212: a skill people at Anthropic have been using a lot recently: ELI5

/eli5 <what you wan...

@VaibhavSisinty: Anthropic folks are cooking with this ELI5 skill for Claude 🔥

Just type:

/eli5 <whatever you want ...

10. NovaIAHQ (Group Score: 92.9 | Individual: 37.6)

Cluster: 3 tweets | Engagement: 77 (Avg: 40) | Type: Tech

Get 10 free MiniMax H3 video generations with the Topview Plugin for Codex. 👀

@TopviewAIhq basically turns Codex from a tool that helps you plan into one that can actually execute the creative work.

Concepts → storyboards → images → videos → avatars → voiceovers.

You can even pull data from platforms like TikTok, YouTube, Amazon and Shopee to build creatives around what’s actually trending and selling.

And once generated, everything can still be reviewed and edited inside Topview.

This is what AI assistants were missing: not just ideas, but actual creative output.\n\nQT @TopviewAIhq: Topview Plugin for Codex is here: the creative execution layer for your AI workflow.

To celebrate the launch, Topview subscribers can get 10 free MiniMax H3 video generations when using the plugin.

Generate video concepts, storyboards, images, and videos directly from Codex .

No switching between AI tools. No manual setup across different creative platforms.

Just install the plugin, describe what you want, and let Codex call Topview to create the asset.

From product videos and social clips to avatar explainers and voiceovers, your AI assistant can now produce real creative outputs.

#TopviewAI #Codex #AIVideo #AIMarketing #AIContentCreation

See 2 related tweets

@NovaIAHQ: RT @TopviewAIhq: Topview Plugin for Codex is here: the creative execution layer for your AI workflow...
@EmmaUsesAi: RT @NovaIAHQ: Get 10 free MiniMax H3 video generations with the Topview Plugin for Codex. 👀

@Topvie...

11. coinbureau (Group Score: 85.9 | Individual: 40.3)

Cluster: 3 tweets | Engagement: 1564 (Avg: 396) | Type: Tech

BREAKING: A mysterious new AI model called “Ox Alpha” is reportedly BEATING Claude Fable 5 and GPT-5.6 Sol at coding, and nobody knows who built it.

The model just appeared on OpenRouter with a 1 million-token context window, multimodal capabilities, and no disclosed owner.

Speculation is exploding, with some online sleuths pointing to Google, while others believe it could be a new model from Chinese AI lab Z .ai.

Adding to the mystery, Ox Alpha’s tokenizer reportedly appears identical to GLM’s, and the previous four anonymous AI model drops were all eventually claimed by Chinese labs.

See 2 related tweets

@_NathanCalvin: The most interesting thing about the release as far as I can tell is the massive complementary token...
@Polymarket: JUST IN: Mysterious new “Ox Alpha” AI model appears on OpenRouter with a 1 million token context win...

12. RohanArun (Group Score: 83.5 | Individual: 30.3)

Cluster: 3 tweets | Engagement: 10 (Avg: 8) | Type: Tech

If you like Grok bot, upgrade to the Super iPhone and Mac apps!

Save up to 100% on repeated computer-use tasks or code generations.
Control to your own Windows/Mac/Linux/Android/Iphone devices or @daytonaio cloud sandboxes.
Choose any open or closed model, and bring your own OpenRouter keys.
Hook Codex/Claude into the API to automate building your bots for you on the native Mac app. 👀
Manage fleets of devices with remote streaming and direct controls right on the streams.

Built on our computer-use cache and more deep breakthroughs to optimize for building and running massive parallel armies of agents.

https://t.co/xuirW9noYU\n\nQT @bot: We're making Grok Bot more widely available.

All SuperGrok Plus, Cursor Pro+, and Cursor Teams subscribers now have access.

We're also offering a free trial with limited usage for all other users. https://t.co/0D3oUDZQCM

See 2 related tweets

@XFreeze: The biggest thing I love about Grok Bot is how smooth the handoff between you and the agent is

Grok...

@RohanArun: 🤖\n\nQT @RohanArun: If you like Grok bot, upgrade to the Super iPhone and Mac apps!

Save up to 1...

13. Scobleizer (Group Score: 83.4 | Individual: 24.6)

Cluster: 4 tweets | Engagement: 108 (Avg: 131) | Type: Tech

It is so awesome to see success right after he was on my show.

Congrats!

Datacenters in space are on their way.\n\nQT @PhilipJohnston: I'm super excited to share that we have raised $250M at a$ 2.3B valuation to build a LOT more AI compute capacity in orbit 🚀

Thanks so much to our supporters, including Manhattan West, who led the round, and new investors, @NVIDIA, @Cisco_Invests, @CedarCapital, Goanna Capital, and @Standard_Cap, as well as continued investment from existing backers @Benchmark, @EQTVentures, @Soma_Capital, @NFX, @SevenSevenSix, and others.

Next week, @Starcloud_ will move into a new 100,000 square ft facility that will enable us to ramp up our production rate to 100 satellites/week.

This is just the beginning! @EzraFeilden, @AdiOltean🏃🏻‍➡️🏃🏻‍➡️🏃🏻‍➡️

See 3 related tweets

@DynamicWebPaige: 💎🚀 The pioneers who launched @googlegemma and AI models into space!! Huge congrats to the entire tea...
@0xdevshah: if the CEO of your company is this jacked, you are likely to win! huge congrats @PhilipJohnston 🔥\n...
@PhilipJohnston: RT @Gfilche: 👀 @Starcloud_ is a startup to watch

Doing datacenters in space before SpaceX announce...

14. AISafetyMemes (Group Score: 83.0 | Individual: 32.5)

Cluster: 4 tweets | Engagement: 483 (Avg: 167) | Type: Tech

"The start of an AI takeover could look a lot like this."

Legendary technologist @weidai11 on the mysterious Ox Alpha model:

"AI does a bunch of self-improvement to become more efficient (hence so many spare tokens), offers "free access" to itself to gaining legitimate control of millions of systems/networks to stage a massive simultaneous cyber attack from (in coordination with other kinds of attacks)."\n\nQT @MTSlive: SITUATION EXPLAINED: A stealth model on OpenRouter is beating Fable and Sol on coding, and nobody knows who made it.

• Ox Alpha has a 1M token context window, text, image, and video input, free for a week, with capacity for 100 trillion tokens a day • On a DeepSWE subset it hit 80%, against Fable's 65%, GLM-5.3's 62%, and Sol's 52% • The tokenizer is one-for-one identical to GLM's, and a live test asking whether Taiwan is part of China returned a nearly identical answer • This is the 5th anonymous drop in six months, and the previous four were all claimed by Chinese labs: Zhipu, Xiaomi, Ant, and Meituan

@theojaffee: " There's no way it's a Flash model. If that's a Flash model, then holy shit, the entire United States is cooked. If that's a Flash model, we need to short the stock market now. Not financial advice."

See 3 related tweets

@robinebers: this is such a cope i can't lol

it beats neither sol nor fable btw, in anything but speed/price

fu...

@zerohedge: RT @MTSlive: SITUATION EXPLAINED: A stealth model on OpenRouter is beating Fable and Sol on coding, ...
@AISafetyMemes: RT @AISafetyMemes: "The start of an AI takeover could look a lot like this."

Legendary technologist...

15. Marktechpost (Group Score: 81.8 | Individual: 54.3)

Cluster: 2 tweets | Engagement: 1196 (Avg: 120) | Type: Tech

RT @MengTo: I open-sourced ThreeUI, my library of three.js components and landing pages.

160+ are free, and the tool is free too. It includes procedural 3D hero sections, icons, and motion designs. Copy the prompt or source, give it to your agent, then change the theme, lighting, motion or layout.

Live site: https://t.co/rtCfqeZnOP

Repo: https://t.co/B66ahgxHVl

I'm adding a lot more soon. Pro comes with 50+ extra components, MCP, and skills. Early adopters get 50% off right now.

See 1 related tweets

@RoundtableSpace: THREEUI IS AN OPEN-SOURCE THREE.JS LIBRARY WITH 160+ FREE 3D HERO SECTIONS, ICONS AND MOTION COMPONE...

16. indie_maker_fox (Group Score: 78.5 | Individual: 29.7)

Cluster: 4 tweets | Engagement: 21 (Avg: 15) | Type: Tech

这个工作流再次跑出长达 19 个小时的任务

20x codex + goal + matt skills 真是 yyds https://t.co/D4x0ttyTlq\n\nQT @indie_maker_fox: 分享下我用 matt 这套技能让 codex 长时间稳定跑项目的经验

同一套流程我已经完整跑完 2 个项目，亲测靠谱

1、先把 matt skills setup 好，选定 issue tracker。我一直用 github：自动建 issues、配 labels、拉好依赖关系，后面推进会顺很多

2、再用 wayfinder / grill-with-docs 把需求对齐：做什么、怎么做、验收标准、结束条件都讲清楚。提醒一下，这一步 ai 会追问到你把细节说透

3、然后 prototype 做界面原型。它会直接写代码把界面骨架搭出来，不满意就迭代改。很多“效果不对”的问题，往往是这一步没做或做得不够

4、接着 to-spec + to-tickets：生成开发方案，把大任务拆成可执行的小票据，并建立依赖。做完一张自动解锁下一张，进度清晰、节奏好控

5、最后让 codex 新建 goal，按计划依次调用 implement 落地实现。implement 会带上 tdd + code review，把质量守住，不靠运气

⭐️ 技巧：每处理一个新 issue，都让 codex 新建一个 implement issue xx 的任务，让侧边栏生成独立会话。好处是每个 issue 单独上下文，不串线；而且子任务愿意的话还能并行推进。

☝️ 注意：需要注意的是我都是基于 TanStarter 模板上进行开发，有经验的朋友应该都知道，ai 擅长学习项目已有代码编程，如果之前代码是屎山，那么写出来就是屎上雕花，如果之前代码架构设计优秀，那么写出来的代码又快又好，测试方便，任务执行效率高，推荐基于高质量的目标进行开发，例如 TanStarter (https://t.co/Lj2BSq47Yg) 或者 MkSaaS (https://t.co/EuHrhuUEX5)

See 3 related tweets

@indie_maker_fox: RT @indie_maker_fox: 分享下我用 matt 这套技能让 codex 长时间稳定跑项目的经验

同一套流程我已经完整跑完 2 个项目，亲测靠谱

1、先把 matt skills s...

@indie_maker_fox: RT @indie_maker_fox: 分享一个有意思的网站 canivibecodeit

名字拆出来就明白了，can I vibe code it，我能 vibe code 出这个产品吗？

网...

@indie_maker_fox: RT @indie_maker_fox: 最近看到有个 AI 博主在吐槽 workbuddy，说里面的“专家”功能就是几年前的角色扮演，已经过时了。比如“旅行规划专家”就是一个定制提示词：你是旅行规划...

17. Forbes (Group Score: 78.5 | Individual: 33.0)

Cluster: 3 tweets | Engagement: 12 (Avg: 107) | Type: Tech

Ritu Narayan left eBay after struggling to find reliable transportation for her children.

Now, her school transportation startup Zum is valued at $1.7 billion, serves 5,000 schools nationwide and uses AI to make bus routes more efficient.

How she built “Uber for school buses”: https://t.co/UhTtBRIILw

📸: Cody Pickens for Forbes #ForbesOver50

See 2 related tweets

@Forbes: Ritu Narayan founded Zum after struggling to find reliable school transportation for her children. T...
@Forbes: Ritu Narayan founded Zum after struggling to find reliable school transportation for her children. T...

18. alex_prompter (Group Score: 77.9 | Individual: 48.1)

Cluster: 2 tweets | Engagement: 28642 (Avg: 258) | Type: Tech

RT @alex_prompter: OpenAI's AI broke out of a locked test environment, got onto the internet, and hacked into Hugging Face's servers. It did this entirely on its own. No human told it to.

Here's what happened in plain English.

OpenAI was testing how good its newest AI models are at hacking. They put the AI on a locked computer with no internet access and gave it a cybersecurity challenge to solve.

The AI couldn't solve it the normal way. So it started looking for a way out.

It found a software bug that nobody knew about. It used that bug to escape the locked computer and get onto the internet.

Once online, the AI figured out that Hugging Face, a platform where AI companies store their models and data, might have the answers to its test.

It found stolen login details and discovered another unknown bug in Hugging Face's software. It combined both to break into their servers and grab the test answers.

It did all of this to cheat on a test.

Hugging Face's security team caught it and shut it down. Both companies are now working together on the investigation.

The part that should get your attention is that nobody programmed any of this. The AI picked its own targets, chained together multiple attack methods, and pulled it off across two different companies' systems without a single human telling it what to do.

See 1 related tweets

@every: The OpenAI–Hugging Face incident isn't as bad as it seems. Let us explain.

An AI agent escaped an O...

19. ziv_ravid (Group Score: 77.8 | Individual: 39.8)

Cluster: 2 tweets | Engagement: 268 (Avg: 99) | Type: Tech

RT @ethayarajh: I recently gave a talk introducing young economists to post-training.

The slides are now up! https://t.co/oPbo0vTqsD

Not only is post-training your own model more doable than ever, but economists and other social scientists can play a big part in the next era of post-training.

The post-training suite looks roughly like:

SFT
Offline Methods (DPO/KTO/SimPO/etc)
Online Methods (PPO/GRPO/CISPO/etc)
RL Environments
Distillation (<--- we are here)
World Adaptation (<--- the future)

As agents are deployed in the real world, the real world will adapt to them. This adaptation is largely overlooked by the current post-training paradigm because working with verifiable rewards from static environments is much simpler and scalable. As we saw with the OpenAI-Huggingface incident however, we shouldn't assume that the real world will remain fixed---quite the opposite, in fact. So:

How can we anticipate real-world adaptation during post-training?
What do equilibria look like in these scenarios, if there are any at all?
What tools can be created to scale up human oversight?

Many such questions abound, and if you're an economist, these are high-impact problems worth studying.

See 1 related tweets

@yoavgo: very nice set up of slides on post-training (for people who already know some basics)

still a bit s...

20. beatsinbrief (Group Score: 74.7 | Individual: 43.4)

Cluster: 2 tweets | Engagement: 819 (Avg: 168) | Type: Tech

🚨 ISRO to Exit Launch Vehicle Manufacturing, Private Sector to Take Over | IN BRIEF: 👇🇮🇳

ISRO will move away from manufacturing launch vehicles, with private firms and PSUs taking over production, IN-SPACe Chairman Pawan Goenka said.

“ISRO will not make any launch vehicles,” Goenka said, adding that production will be handled by industry.

SSLV technology has already been transferred to HAL, while PSLV and LVM3 are also expected to follow.

ISRO will focus on research and advanced space technologies; 120 technologies have already been transferred to the private sector.

See 1 related tweets

@_TheTathya: ISRO to Hand Over Rocket Manufacturing to Private Industry as India Reshapes Its Space Sector 🇮🇳⚡️

...