科技推文精选 - 2026-02-25

2026年2月25日科技每日简报

Today's top tech conversations are led by @starbuxman, whose post about 'RT @ns123abc: 🚨 BREAKING: IBM ...' garnered the highest engagement. Key themes trending across the top stories include engineering, cerebras, massive, their, https. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.

1. starbuxman (Group Score: 137.8 | Individual: 37.5)

Cluster: 6 tweets | Engagement: 1802 (Avg: 243) | Type: Tech

RT @ns123abc: 🚨 BREAKING: IBM stock down 13% after Anthropic announced that Claude can streamline COBOL code

IBM’s entire business model:…

See 5 related tweets

@aakashgupta: Anthropic just vaporized $31 billion of IBM’s market cap with a blog post about COBOL.

The math loo...

@KirkDBorne: IBM stock gets slammed, becoming the latest perceived victim of rapidly developing AI technology, af...
@Cointelegraph: 🚨 UPDATE: IBM drops 13% losing $31B in value as Anthropic claims Claude can read and modernize legac...
@mariofusco: It makes totally sense that yesterday IBM stocks had a 13% fall because Anthropic said that they can...
@minchoi: RT @minchoi: It's over for legacy tech consulting..

IBM dropped ~13% yesterday after Anthropic said...

2. GenAI_is_real (Group Score: 96.3 | Individual: 34.3)

Cluster: 3 tweets | Engagement: 162 (Avg: 58) | Type: Tech

ibm losing $40b because claude can read cobol is a massive overreaction. current models have always understood cobol; the real bottleneck isn't the "translation"—it's the secure, on-prem execution.

no major bank is sending their core financial logic to a public api. the real alpha in 2026 isn't just "reading" old code; it's building serving infra like sglang that's lean enough to run these modernization agents locally behind a firewall.

ibm’s moat isn't the language, it’s the trust. but if we can make on-prem inference 10x more efficient, that moat actually starts to evaporate.

See 2 related tweets

@GenAI_is_real: the ibm crash is the first real market pricing of "legacy technical debt." for decades, the moat was...
@GenAI_is_real: the $ibm crash is the first real market pricing of "obscurity as a moat."

for decades, ibm’s value ...

3. svpino (Group Score: 89.0 | Individual: 48.4)

Cluster: 3 tweets | Engagement: 12046 (Avg: 517) | Type: Tech

RT @AnthropicAI: We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax.

These labs cr…

See 2 related tweets

@HarperSCarroll: MODEL DISTILLATION, EXPL–AI–NED

Anthropic just announced that they identified "industrial-scale dis...

@FT: Anthropic accuses Chinese AI labs of ‘distillation attacks’ on its models https://t.co/gHkuLdZg0J...

4. BrianRoemmele (Group Score: 78.3 | Individual: 47.0)

Cluster: 3 tweets | Engagement: 17200 (Avg: 533) | Type: Tech

RT @elonmusk: Anthropic is guilty of stealing training data at massive scale and has had to pay multi-billion dollar settlements for their…

See 2 related tweets

@MarioNawfal: Anthropic stole training data on a massive scale and had to pay multi-billion dollar settlements for...
@MarioNawfal: RT @MarioNawfal: Anthropic stole training data on a massive scale and had to pay multi-billion dolla...

5. KirkDBorne (Group Score: 76.1 | Individual: 22.7)

Cluster: 4 tweets | Engagement: 146 (Avg: 51) | Type: Tech

Download 674-page PDF >> Introduction to Machine Learning (textbook on foundations, algorithms, and techniques): https://t.co/deN091mk8B ———— #MachineLearning #ML #DeepLearning #AI #Mathematics #DataScience #DataScientist https://t.co/eEykfXbbI7

See 3 related tweets

@KirkDBorne: (Download 462-page PDF eBook)

#DataScience Theories, Models, Algorithms, and Analytics: https://t.c...

@KirkDBorne: Download 284-page PDF “Introduction to #NeuralNetworks” ➡️ https://t.co/0nNNi7ezRu ————— #DataScienc...
@KirkDBorne: Mathematical Theory of #DeepLearning [download 333-page PDF]: https://t.co/T9ydk2MVKa UPDATED

...

6. rohanpaul_ai (Group Score: 70.9 | Individual: 36.4)

Cluster: 3 tweets | Engagement: 182 (Avg: 113) | Type: Tech

The super-viral blog by Citrini, 'THE 2028 GLOBAL INTELLIGENCE CRISIS' - a must-read for real. 🎯

The big idea is this: for modern history, human intelligence was the most scarce and valuable economic resource.

If society makes intelligence cheap and abundant, human labor loses its premium value. Because the entire global financial system was explicitly built around humans earning high wages, abundant AI does not just disrupt jobs, it fundamentally breaks the foundation of the economy.

The authors see AI capabilities accelerating so fast that companies are firing highly paid white collar workers and replacing them with autonomous agents. This creates a devastating, unstoppable loop.
The displaced professionals take massive pay cuts or move to gig work, so they stop buying extra goods. Consumer businesses lose revenue, so they fire more people and buy more AI to protect their margins, which only accelerates the job losses.
At the same time, personal AI agents destroy middleman businesses. Because AI does not get lazy, it instantly cancels unused subscriptions, auto negotiates bills, and bypasses credit card networks to route payments for a fraction of a penny. Entire industries built on consumer friction and brand loyalty collapse.
This breaks the private credit market. Billions of dollars were loaned to software companies assuming their subscription revenues were permanent. When AI agents easily replicate these software tools, the revenues vanish, causing massive loan defaults that threaten the life insurance and retirement funds backing them.
The housing market follows. The $13 trillion mortgage system is built on the assumption that prime borrowers will maintain their high salaries for decades. Unlike the 2008 housing crash, these were fundamentally good loans made to people with excellent credit. But as AI permanently destroys earning power, perfectly underwritten mortgages start defaulting.
Even the government breaks. Federal tax revenue relies heavily on taxing human wages. As AI shifts wealth from human labor to the owners of computer servers, income tax receipts plunge right at the exact moment millions of displaced citizens need government bailouts.

See 2 related tweets

@kimmonismus: One year left:

„We believe it is plausible, as soon as early 2027, that our AI systems could fully ...

@AISafetyMemes: RT @AISafetyMemes: THE 2028 GREAT INTELLIGENCE CRISIS - SUMMARY

AI got good at coding → compani...

7. testingcatalog (Group Score: 68.4 | Individual: 28.6)

Cluster: 4 tweets | Engagement: 370 (Avg: 319) | Type: Tech

BREAKING 🚨: ProducerAI, a tool that helps users to generate music videos, is joining Google Labs.

“ProducerAI is a creative collaborator, whether you’re writing lyrics, refining a melody or inventing entirely new genres.”

A music agent 👀 https://t.co/g0mOvmaJ99

See 3 related tweets

@gdgtify: RT @GoogleLabs: 🚨 Big news: @producer_ai is officially joining Google Labs! 🎶

ProducerAI is a crea...

@TechCrunch: Music generator ProducerAI joins Google Labs https://t.co/k7qWg1aSMW...
@GoogleDeepMind: RT @GoogleLabs: 🚨 Big news: @producer_ai is officially joining Google Labs! 🎶

ProducerAI is a creat...

8. Shashikant86 (Group Score: 65.4 | Individual: 26.8)

Cluster: 3 tweets | Engagement: 4 (Avg: 2) | Type: Tech

🎙️ Announcing Agent Engineering Conference (AgentEng) 2026: A Highly Technical Conference for Agent Engineering is happening in London 🇬🇧 and San Francisco 🌉 ⚡️ 💡For the past few months, I’ve been quietly working on something different in Agentic AI space, thinking of something highly curated, highly technical, demo-first conference on Agentic Engineering where speakers showcase code, IDEs, CLIs, Demos and not just slides.

🤖Prompt Engineering → Context Engineering → Harness Engineering → Eval Engineering→ → Skills Engineering → Memory Engineering → All roads lead to Agent Engineering So we’re launching AgentEng 2026. 📍 London · Summer 2026 🌉 San Francisco edition to follow Single track. Demo-first sessions. Built on the 2,000+ member London Agentic AI community and strong connections across Silicon Valley. 🎤 CFP is now open. ⚡️Founding partners invited. Checkout : https://t.co/CAJhkX9kwG

See 2 related tweets

@Shashikant86: Highly inspired by the work from the @swyx on @aiDotEngineer Conference. Agent Engineering Conferenc...
@Shashikant86: One company that talked about Agent Engineering a lot is @LangChain and @LangChain_OSS. You should d...

9. rohanpaul_ai (Group Score: 65.0 | Individual: 33.3)

Cluster: 2 tweets | Engagement: 39 (Avg: 113) | Type: Tech

Cerebras Systems is confidentially filing for a US IPO targeted for the 2nd-Quarter of 2026.

This aggressive comeback follows the withdrawal of their previous filing in late 2025, which faced regulatory hurdles due to heavy reliance on UAE-based investor G42.

Cerebras secured massive private funding, pushing its reported valuation to $23 billion.

A major hurdle for the first IPO attempt was customer concentration. Cerebras needed a massive, U.S.-based client to satisfy both regulators and skeptical investors.

This is where the transformative $10B, multi-year deal with OpenAI comes into play in January, 2026

OpenAI is securing 750 megawatts of compute capacity through 2028 via Cerebras' cloud services. This focuses strictly on inference (running the AI models), leveraging Cerebras' massive wafer-scale chips to generate responses up to 15 times faster than standard GPUs.

OpenAI CEO Sam Altman was actually an early personal investor in Cerebras. The two companies have maintained regular discussions since 2017, culminating in this massive contract after Cerebras proved its hardware efficiency last year.

For OpenAI, this deal with Cerebras significantly reduced their reliance on Nvidia. For Cerebras, this perfectly diversifies its revenue away from G42 right before their new public listing.

See 1 related tweets

@aakashgupta: Sam Altman is personally invested in Cerebras. Then OpenAI signed a $10B deal with Cerebras. Now Cer...

10. Reuters (Group Score: 64.5 | Individual: 27.5)

Cluster: 3 tweets | Engagement: 531 (Avg: 145) | Type: Tech

Exclusive: China's DeepSeek trained AI model on Nvidia's best chip despite US ban, official says https://t.co/bA3HuNSP8i https://t.co/bA3HuNSP8i

See 2 related tweets

@bindureddy: Lots of allegations about how DeepSeek has trained their models
they distilled both OpenAI and A...
@MarioNawfal: 🚨🇨🇳 DeepSeek trained their new AI model on Nvidia's most advanced AI chip. They're also banned from ...

11. HamelHusain (Group Score: 61.7 | Individual: 39.5)

Cluster: 3 tweets | Engagement: 1566 (Avg: 174) | Type: Tech

RT @noahzweben: Announcing a new Claude Code feature: Remote Control. It's rolling out now to Max users in research preview. Try it with /r…

See 2 related tweets

@testingcatalog: Anthropic is rolling out a new Remote Control feature that allows users to pick up Claude Code tasks...
@rohanpaul_ai: RT @Meer_AIIT: Just released by Anthropic: Remote Control for Claude Code.

Start a task in your ter...

12. arimorcos (Group Score: 60.4 | Individual: 29.5)

Cluster: 3 tweets | Engagement: 434 (Avg: 65) | Type: Tech

RT @StefanoErmon: Mercury 2 is live 🚀🚀

The world’s first reasoning diffusion LLM, delivering 5x faster performance than leading speed-opti…

See 2 related tweets

@testingcatalog: BREAKING 🚨: Inception has launched Mercury 2, the first reasoning diffusion LLM with 5x the performa...
@mark_k: Mercury 2 was launched by @_inception_ai, the world's first reasoning diffusion LLM. 🗣️

With 5x fas...

13. karpathy (Group Score: 56.7 | Individual: 33.4)

Cluster: 2 tweets | Engagement: 6671 (Avg: 11440) | Type: Tech

CLIs are super exciting precisely because they are a "legacy" technology, which means AI agents can natively and easily use them, combine them, interact with them via the entire terminal toolkit.

E.g ask your Claude/Codex agent to install this new Polymarket CLI and ask for any arbitrary dashboards or interfaces or logic. The agents will build it for you. Install the Github CLI too and you can ask them to navigate the repo, see issues, PRs, discussions, even the code itself.

Example: Claude built this terminal dashboard in ~3 minutes, of the highest volume polymarkets and the 24hr change. Or you can make it a web app or whatever you want. Even more powerful when you use it as a module of bigger pipelines.

If you have any kind of product or service think: can agents access and use them?

are your legacy docs (for humans) at least exportable in markdown?
have you written Skills for your product?
can your product/service be usable via CLI? Or MCP?
...

It's 2026. Build. For. Agents.

See 1 related tweets

@nummanali: Karpathy Chronicles

He believes CLIs are a better interface for agents
He wrote this post on...

14. rohanpaul_ai (Group Score: 55.2 | Individual: 33.5)

Cluster: 2 tweets | Engagement: 394 (Avg: 113) | Type: Tech

RT @rohanpaul_ai: "If you really want to make money, found an agentic AI company.

I mean, build an agent to do something. This is the ag…

See 1 related tweets

@eladgil: Will make AI agents via @dreamer for you this AM w/ @dps and @swyx

Post a reply in-thread @dreamer...

15. KirkDBorne (Group Score: 52.0 | Individual: 25.3)

Cluster: 3 tweets | Engagement: 39 (Avg: 51) | Type: Tech

Python for Geospatial Data Analysis: https://t.co/klUrKwmZc0
Guide to Geospatial Analysis with SQL: https://t.co/z1sqGv0HYS

————— #DataScience #Geoscience #LocationIntelligence #GIS #Geoinformatics #ML #MachineLearning #DataScientist https://t.co/IVR1sP1wbf

See 2 related tweets

@KirkDBorne: Learning #Geospatial Analysis with #Python — Unleash the power of Python with practical techniques f...
@KirkDBorne: Applied Geospatial #DataScience with Python: https://t.co/8tJooRlwHA v/ @PacktDataML ————— #GIS #Lo...

16. ReutersBiz (Group Score: 50.8 | Individual: 17.1)

Cluster: 3 tweets | Engagement: 42 (Avg: 15) | Type: Tech

WATCH: Advanced Micro Devices agreed to sell up to $60 billion worth of artificial-intelligence chips to Meta Platforms over five years in a deal that allows the Facebook owner to purchase as much as 10% of the chip firm https://t.co/qJNnAXz7g6 https://t.co/WQyaOwchea

See 2 related tweets

@WSJ: Meta struck an AI-chip deal with Advanced Micro Devices worth more than $100 billion https://t.co/Go...
@Reuters: Advanced Micro Devices agreed to sell up to $60 billion worth of artificial-intelligence chips to Me...

17. MarioNawfal (Group Score: 49.1 | Individual: 30.7)

Cluster: 2 tweets | Engagement: 6008 (Avg: 1347) | Type: Tech

🚨🇺🇸 Breaking: xAI just signed a deal with the U.S. Department of Defense to bring Grok into classified systems.

The Pentagon’s bringing in the real one.

When it comes to cutting-edge AI for the most sensitive stuff, they chose Grok.

Smart move.

Source: Axios https://t.co/9tbrIIPhmz

See 1 related tweets

@MarioNawfal: 🚨🇺🇸 Elon's xAI just signed a deal to bring Grok into the U.S. military's classified systems, making ...

18. Franc0Fernand0 (Group Score: 48.7 | Individual: 31.7)

Cluster: 2 tweets | Engagement: 196 (Avg: 77) | Type: Tech

Shipping faster doesn’t mean shipping better.

Any seasoned engineer can tell you that success in software isn’t about pumping out more lines of code.

It’s about building systems that are safe, can be maintained, are observable, and long-lasting.

See 1 related tweets

@Franc0Fernand0: RT @Franc0Fernand0: Shipping faster doesn’t mean shipping better.

Any seasoned engineer can tell yo...

19. garrytan (Group Score: 47.1 | Individual: 24.2)

Cluster: 2 tweets | Engagement: 639 (Avg: 312) | Type: Tech

AI will unleash the most vibrant new market there could be. Cognition becoming cheap means we will do more with less, emphasis on doing more. https://t.co/jGKgs06jTi

See 1 related tweets

@burkeholland: I have considered that the real value of AI might not be AI at all.

It might force us to figure ou...

20. llama_index (Group Score: 46.4 | Individual: 33.9)

Cluster: 3 tweets | Engagement: 48 (Avg: 35) | Type: Tech

Document OCR benchmarks are hitting a ceiling - and that's a problem for real-world AI applications.

Our latest analysis reveals why OmniDocBench, the go-to standard for document parsing evaluation, is becoming inadequate as models like GLM-OCR @Zai_org achieve 94.6% accuracy while still failing on complex real-world documents.

📊 Models are saturating OmniDocBench scores but still struggle with complex financial reports, legal filings, and domain-specific documents 🎯 Rigid exact-match evaluation penalizes semantically correct outputs that differ in formatting (HTML vs markdown, spacing, etc.) ⚡ AI agents need semantic correctness, not perfect formatting matches - current benchmarks miss this critical distinction 🔬 The benchmark's 1,355 pages can't capture the full complexity of production document processing needs

The document parsing challenge isn't solved just because benchmark scores look impressive. We need evaluation methods that reward semantic understanding over exact formatting, especially as AI agents become the primary consumers of parsed content.

We're building parsing models focused on semantic correctness for complex visual documents. If you're scaling OCR workloads in production, LlamaParse handles the edge cases that benchmarks miss.

Read our full analysis: https://t.co/tcZP1PM8kv

See 2 related tweets

@jerryjliu0: RT @llama_index: Document OCR benchmarks are hitting a ceiling - and that's a problem for real-world...
@itsclelia: RT @jerryjliu0: OmniDocBench is getting saturated

VLMs are getting increasingly better at document...