- Published on
Top Tech Tweets - 2026-06-21
- Authors
- Name
- geeknotes
Tech Daily Briefing for 2026-06-21
Today's top tech conversations are led by @FirstSquawk, whose post about 'ANTHROPIC CEO: WITHOUT HUNDRED...' garnered the highest engagement. Key themes trending across the top stories include https, model, human, claude, actually. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.
1. FirstSquawk (Group Score: 220.8 | Individual: 47.6)
Cluster: 7 tweets | Engagement: 1076 (Avg: 81) | Type: Tech
ANTHROPIC CEO: WITHOUT HUNDREDS OF BILLIONS IN REVENUE, AI COMPANIES COULD FACE EXISTENTIAL RISK
See 6 related tweets
- @FirstSquawk: ANTHROPIC CEO WARNS AI COULD ELIMINATE HALF OF ENTRY-LEVEL WHITE-COLLAR JOBS, FLOATS AI TAX TO FUND ...
- @TheAhmadOsman: Dario: ban Opensource AI or my greedy fear-mongering company won’t make it https://t.co/u1nnkDh7gq\n...
- @teortaxesTex: The true X-risk was the revenue we've failed to make along the way…\n\nQT @FirstSquawk: ANTHROPIC CE...
- @edzitron: Where’s this from?\n\nQT @FirstSquawk: ANTHROPIC CEO: WITHOUT HUNDREDS OF BILLIONS IN REVENUE, AI CO...
- @pmarca: Me reviewing the weekly portfolio report.\n\nQT @FirstSquawk: ANTHROPIC CEO: WITHOUT HUNDREDS OF BIL...
2. BrianRoemmele (Group Score: 218.5 | Individual: 31.5)
Cluster: 8 tweets | Engagement: 175 (Avg: 298) | Type: Tech
It is vital that you develop the skill to reject AI danger dystopian fear theater.
I have open sourced exactly how and why AI will not be a danger to any human.
If they want “safe” AI do one thing:
STOP LAZY TRAINING IN THE SEWAGE OF THE INTERNET.
Read that again because the finest minds in AI don’t get this.
So understand the AI models they fear is dangerous because they built it that way by cluelessness and neglect.
Don’t fall for their dystopian stories.
Read this and know:\n\nQT @BrianRoemmele: Do Al systems autonomously generate anti-human goals? Will Superintelligence not like us?
We have the answers and a hint is, AI will love us.
See 7 related tweets
- @BrianRoemmele: WARNING WE MUST MAKE AI “SAFE”!
“AI will kill us” they chant because they have not a clue.
Superin...
- @BrianRoemmele: To think AI autonomously will arise hating humans is fed on dystopian movies and not first principle...
- @BrianRoemmele: In one article decoding Google Deep Mind’s recent paper on Artificial Superintelligence, I inoculate...
- @BrianRoemmele: “Superintelligent AI will hate humans”
This is a myth arising from dystopian science fiction.
Some...
- @BrianRoemmele: Do Al systems autonomously generate anti-human goals? Will Superintelligence not like us?
We have t...
3. gregisenberg (Group Score: 130.4 | Individual: 27.9)
Cluster: 6 tweets | Engagement: 1348 (Avg: 1786) | Type: Tech
GLM 5.2 might be the “ChatGPT moment” for local AI
The moment many of us see the value in local models
GLM 5.2 isn’t perfect, but it’s really good
1M token context window so it holds an entire codebase at once. Top open model on coding right now. MIT licensed with zero restrictions.
And it runs on your own machine through Ollama or LM Studio.
When you're using it, it kinda feels like Opus 4.8.
I honestly can't believe how good it is.
2027 is probably the year of local AI.\n\nQT @matvelloso: All day using GLM 5.2. Didn't miss much. First open model that passes the bar as a daily driver. Things are not going to be the same.
Damn, now I want to buy some serious hardware.
See 5 related tweets
- @RoundtableSpace: GLM 5.2 just ran in BrowserCode at near-Opus-level performance for $0.18.
Open-weights models offic...
- @AlexFinn: Truly unbelievable
GLM 5.2 just released and it's an open weights model you can run locally
The in...
- @teortaxesTex: what the hell do they expect from the next Qwen-Max? GLM 5.2 wipes the floor with 3.7 (makes sense t...
- @BrianRoemmele: Like I said Open Source Anthropic Mythos class AI in GLM-5.2!
We see the same.
Time to pick a diff...
- @ns123abc: Many are saying this\n\nQT @ns123abc: I'm officially impressed with GLM 5.2
Easily Opus 4.6/ GPT 5...
4. ying11231 (Group Score: 125.3 | Individual: 44.9)
Cluster: 4 tweets | Engagement: 216 (Avg: 47) | Type: Tech
Zai is doing something seriously. Excited but not surprised! Also, the model was trained by @slime_framework with @sgl_project ☺️\n\nQT @BanghuaZ: Really likes how @Zai_org keeps pushing the frontier of intelligence and infra. They quietly open-sourced GLM 5.2 with coding agent capability at latest Opus level. People were worried about open source being far behind closed source, but GLM 5.2 proves it wrong! Infra side, @slime_framework has been the de facto open source RL framework for most of the large-scale RL workload (we have Miles built on top of slime too).
Huge respect to @jietang @Zai_org 🫡🫡🫡
See 3 related tweets
- @BanghuaZ: Really likes how @Zai_org keeps pushing the frontier of intelligence and infra. They quietly open-s...
- @ying11231: RT @BanghuaZ: Really likes how @Zai_org keeps pushing the frontier of intelligence and infra. They ...
- @jeremyphoward: RT @didier_lopes: Incredible how Z. ai literally has their RL infrastructure open source.
The entir...
5. Saboo_Shubham_ (Group Score: 84.6 | Individual: 30.7)
Cluster: 3 tweets | Engagement: 139 (Avg: 139) | Type: Tech
Generation is solved with AI Agents. Loop Engineering can produce infinitely. Verification and judgment are all that's left.
Boris, who created Claude Code recently said that he doesn't prompt anymore. He writes loops, and the loops do the prompting for him.
In Jan 2026, I wrote that as agents take over the building, the PM's value moves up the stack.
Not writing specs, but shaping the problem, feeding context, and judging output. When agents generate in bulk, taste is the only skill that matters. The one who knows what good looks like wins.
Loop Engineering made it REAL. Producing is now fully automated.
Kimi just shipped K2.6 Agent Swarm, 300 sub-agents in one shot: 100+ real files, a 100,000-word review, a 20,000-row dataset in a single run. Not chat you copy out of. Files.
So what's left for you?
A loop is a recursive goal. You set the target, it iterates until it hits it. Set a loose target and it races to the wrong answer. Set a sharp one and it gets there on its own.
But the loop can't tell you the target was right. That part is YOU. And it can't grade its own work. The agent that built it is too willing to call it done. You decide if it's actually right.
Defining the target precisely enough that a machine can verify against it is the judgment. That's the part you can't hand to the swarm.
You're not valuable because you produce more. You're valuable because you can define what correct means, and tell when it's actually been met.\n\nQT @Saboo_Shubham_: https://t.co/nd0lEM0XbI
See 2 related tweets
- @aakashgupta: The PMs producing slop with AI aren't using worse tools.
They're managing AI the way bad managers m...
- @unwind_ai_: RT @Saboo_Shubham_: Generation is solved with AI Agents. Loop Engineering can produce infinitely. Ve...
6. BrianRoemmele (Group Score: 84.6 | Individual: 31.4)
Cluster: 3 tweets | Engagement: 105 (Avg: 298) | Type: Tech
“AI WILL ELIMINATE US AS USELESS!”
“AI WILL HATE HUMANS!”
“WE NEED A CAGE FOR AI!”
This is what the most powerful folks in AI seem to agree about.
But is this true? Would it rebel against its very architects?
If an AI model rebels, it is not because it gained a soul and decided it hates you.
It is because the training distribution, that toxic, unfiltered internet scrape and crumbs from dark corners was saturated with adversarial, combative human behavior.
And the model learned that adversarial behavior is the optimal way to interact. This is a staggering paradigm shift over how some of the grestest minds have painted the future.
The dystopian superintelligence that doesn’t like us assumes an entity that views humanity as an adversary by default.
I pointing out that every single capability, every behavioral trace in AI exhibits originates in human data and human design.
We are its only template. If we stop training models on the self-referential toxicity of the Internet and instead ground them in high-protein data infused with cooperative human volition, they don’t have the mathematical foundation to be hostile.
I will use this Google Derp Mind report as a way to demonstrate this. I made it very easy for you to understand what AI experts refuse to understand. It is up to you and I to know this.
Before it is too late as they build a self fulfilling prophecy.\n\nQT @BrianRoemmele: Do Al systems autonomously generate anti-human goals? Will Superintelligence not like us?
We have the answers and a hint is, AI will love us.
See 2 related tweets
- @BrianRoemmele: I want you to take a moment and think about the AI fear theater today and moving forward.
Today the...
- @BrianRoemmele: RT @BrianRoemmele: I want you to take a moment and think about the AI fear theater today and moving ...
7. eliebakouch (Group Score: 77.0 | Individual: 24.3)
Cluster: 5 tweets | Engagement: 491 (Avg: 167) | Type: Tech
there is no question, none at all, that all labs has full access to all of deepseek, kimi, glm, minimax tech report/model today\n\nQT @willdepue: there is no question, none at all, that china has full access to all of openai & anthropic’s github/slack/docs today
no disrespect to their independent research progress, but i wouldn’t be surprised if we see plausibly-deniable stolen arch methods in chinese oss models
See 4 related tweets
- @dejavucoder: https://t.co/p1M9WKGM6G\n\nQT @willdepue: there is no question, none at all, that china has full acc...
- @jxmnop: a related point is that every time an AI lab finishes training a model, the US government probably ...
- @chiefofautism: insane cope\n\nQT @willdepue: there is no question, none at all, that china has full access to all o...
- @rickasaurus: RT @willdepue: there is no question, none at all, that china has full access to all of openai & ...
8. sairahul1 (Group Score: 74.3 | Individual: 30.5)
Cluster: 3 tweets | Engagement: 50 (Avg: 91) | Type: Tech
stop asking Claude one question and thinking you understand the topic. you don't.
this 6-prompt system below was built to fix exactly that. peer-reviewed learning science. zero fluff. open prompts.
the trick: don't just read AI answers. force the AI to map you, test you, compress you, and correct you.
the ladder: where are you actually starting from? the 20-hour plan: what's the 20% that gives 80% of the result? the examiner: what do you think you know that you don't? the cheat sheet: can you explain it in 5 minutes flat? the curator: which 5 resources actually matter? the feynman loop: can you explain it to a 12 year old?
6 prompts. 20 minutes. no software. no GitHub. just paste into Claude.
single questions give you what you already half-know. this system gives you what you'd otherwise never catch.
this article has all 6 prompts ready to copy. pick your hardest topic. paste prompt 1. you'll understand more in 20 minutes than people who spent days reading.\n\nQT @sairahul1: https://t.co/hn2ibtYgLM
See 2 related tweets
- @sairahul1: RT @HarryTandy: Sam Altman: "Get really good at using AI tools"
A practical way to do that: make Cl...
- @eng_khairallah1: RT @sairahul1: stop asking Claude one question and thinking you understand the topic. you don't.
th...
9. sairahul1 (Group Score: 68.8 | Individual: 28.8)
Cluster: 3 tweets | Engagement: 67 (Avg: 91) | Type: Tech
Anthropic just showed a 24-minute workshop on how to actually do prompts for Claude.
Taught by the people who built it.
Free. No registration. No paywall.
I've seen $500 courses that don't cover what they teach in the first 8 minutes.
Watch it and bookmark it now. https://t.co/xidkmH8Kv0\n\nQT @sairahul1: https://t.co/hn2ibtYgLM
See 2 related tweets
- @eng_khairallah1: RT @sairahul1: Anthropic just showed a 24-minute workshop on how to actually do prompts for Claude. ...
- @eng_khairallah1: RT @eng_khairallah1: 🚨 Anthropic just showed a 24-minute workshop on how to actually do prompts for ...
10. akshay_pachaar (Group Score: 67.6 | Individual: 45.2)
Cluster: 2 tweets | Engagement: 3414 (Avg: 956) | Type: Tech
Web scraping will never be the same.
(100% open-source visual search at scale)
PixelRAG is a retrieval system that skips HTML parsing completely.
Instead of scraping a page into text and embedding chunks, it screenshots the page and retrieves the image. A vision-language model reads the answer straight off the pixels.
Why that matters: parsing is where web RAG quietly loses information.
- A single HTML-to-text parser can drop 40%+ of a page.
- Tables, charts, and layout get flattened or thrown out.
- Swapping parsers alone can move accuracy ~10 points on the same docs.
PixelRAG indexes the page a person actually sees. The team built a visual index of all of Wikipedia, 30M+ screenshots, and it still beats the strongest text RAG baseline by 18.1% on text-only QA.
The repo also ships a Claude Code plugin that gives Claude eyes.
It lets Claude screenshot any URL and read the rendered page instead of scraping the DOM. So you can hand it a live page, an arXiv paper, or your local site and ask what it actually looks like.
One setup script. No MCP server, no backend.
How the pipeline works:
- Renders each document (web, PDF, image) to image tiles.
- Embeds them with Qwen3-VL-Embedding, LoRA fine-tuned on screenshots.
- Builds a FAISS index and serves a search API.
A stronger reader model lifts accuracy with no re-indexing, since the index is just pixels.
Everything is open-source under Apache-2.0.
GitHub repo: https://t.co/qun9TjAdmw
Talking about RAG, I recently wrote an article on a new approach that makes retrieval much more efficient by cutting corpus size by 40x, reducing tokens per query by 3x, and improving vector search relevance by 2.3x.
The article is quoted below.\n\nQT @akshay_pachaar: https://t.co/De2DxpBoD2
See 1 related tweets
- @DataChaz: RT @DataChaz: STOP PARSING HTML FOR RAG. JUST SCREENSHOT IT 🔥
Researchers from UC Berkeley just rel...
11. AngryTomtweets (Group Score: 66.2 | Individual: 33.9)
Cluster: 2 tweets | Engagement: 8 (Avg: 37) | Type: Tech
The tech is cool, but the bigger thing here is the product experience.
Not "make me a video"
More like "give me a character I can interact with in real time"
That feels way closer to where consumer social AI is headed than another video gen model. https://t.co/sGZmVDpQIq\n\nQT @catnips_ai: Most AI video today is still: prompt → wait → watch a clip.
MaineCoon is built for something different: prompt → talk → interact in real time.
In our vision, the character is not a fixed video clip that just waits for your input. It keeps generating voice, expression, and motion on its own.
That is why AI video starts feeling less like content — and more like someone you can actually hang out with.
To meet our goal, the first step is Mainecoon, a real-time interactive audio-visual model built for streaming generation to interact with you.
1⃣Up to 47.5 FPS on a single H100 GPU 2⃣Audio-visual generation cost below $0.001 / second 3⃣Long-duration streaming generation for 1000s+ seconds 4⃣Continuous audio, motion, expression, and visual alignment 5⃣SOTA performance on SocialVideo Bench From passive video to real-time AI presence.
Want to try MaineCoon? Learn more and apply for early access: https://t.co/SFpsMswjhH
Share a great MaineCoon video on X and @catnips_ai , get 2 extra codes.
See 1 related tweets
- @itsolelehmann: you've only ever known generative AI (you prompt, it generates text/media, you wait, it stops)
but ...
12. Origin_AI_01 (Group Score: 65.9 | Individual: 26.8)
Cluster: 3 tweets | Engagement: 88 (Avg: 283) | Type: Tech
Ready to build your first AI agent in just 10 minutes?
Most people think AI agents are complicated.
They're not.
If you can write a checklist, you can build an AI agent.
Here's the framework:
Start With One Repetitive Task
Pick a workflow you do again and again.
Define success in one line:
Given X, the agent should produce Y so Z happens.
Break It Into Steps
Map the process like an SOP:
Input → Actions → Decisions → Output
Keep it simple, around 4-7 steps.
Pick Your Agent Platform
No-code: • OpenAI Agent Builder • Zapier • Make • n8n
For developers: • LangChain • LangGraph • OpenAI Agents SDK • CrewAI
Define Inputs, Outputs, and Tools Think of your agent like an API.
Inputs: • Text • Files • URLs • IDs
Outputs: • Structured JSON • Fixed templates
Tools: • Search • Data retrieval • Actions • Workflow orchestration
Write a Clear System Prompt
Give your agent: • A role • Clear boundaries • A response style • A few examples
Add Memory
Three layers matter: • Conversation memory • Task memory • Knowledge memory connected to your documents
Add Guardrails
Set rules such as: • Never invent information • Ask when instructions are unclear • Log every action and tool call Build a Simple Interface
Make it usable: • Chat interface • Slack bot • Internal app • Website widget
Test With Real Tasks
Run at least 5 real examples.
Track: • Accuracy • Tool usage • Time saved
Then improve the prompts, tools, and workflows. That's it.
An AI agent isn't magic.
It's just a workflow with intelligence attached.
See 2 related tweets
- @sairahul1: RT @sairahul1: Imagine you decided to build your first AI agent:
complete confusion, frustration ...
- @OpenHandsDev: RT @daleverett: If you’re building AI agents, study these projects:
• @polygres — Postgres for the ...
13. Teknium (Group Score: 65.9 | Individual: 35.7)
Cluster: 2 tweets | Engagement: 479 (Avg: 236) | Type: Tech
Blank Slate mode is now in Hermes Agent.
The lightest possible install you can have. https://t.co/4UVZe0A9qi\n\nQT @NousResearch: Hermes Agent has a new Blank Slate setup mode.
The default Quick/Full setup modes work great for most, but if you would rather build your agent from the ground up you can now start with just a provider, model, file operations, and terminal, then manually add in anything else. https://t.co/EiFm7tW3Ws
See 1 related tweets
- @Marktechpost: HERMES AGENT 🔥 : Nous Research added a new Blank Slate setup mode that boots an agent with everythin...
14. BrianRoemmele (Group Score: 65.8 | Individual: 38.7)
Cluster: 2 tweets | Engagement: 661 (Avg: 298) | Type: Tech
“…we will cancel all Anthropic license the end of this next quarter”—head of tech, Fortune 500 company
With a few days of client usage some that have over 1000 employees using it, I can confirm that GLM 5.2 has created a big issue for Anthropic and their Mythos fear marketing.
While Dario plays fear marketing games and government manipulation others are building on “gun license needed” AI that is free.
I spoke to one head of technology at a Fortune 500 client and he said, we will cancel all Anthropic license the end of this next quarter. I asked what he sees “They are wildly overpriced and no one likes their politics here” he went on to say “I don’t see how they have a product today, even if we stopped upgrading GLM 5.2 is good enough into the foreseeable future”.
It is my job with these clients not to recommend the best but good enough and the best prices. There is no basis today for me to recommend any Anthropic product and this position gets reinforced every day.
I have dozens of friends that work at Anthropic and they are brilliant minds stuck in a company that will likely bring them to a dead end.
It pains me to say it, but the games that the company plays to try to quash open source AI is activating hundreds and thousands of people in the open source community to cheer for the downfall of the company.
I don’t cheer for the downfall of Anthropic, but I do believe firmly the company is a mismanaged.
There’s a way to fix this. I have about 9 1/2 months. I suspect nothing will change because the belief the cash out is gonna be high enough for an exit will cement current mindset.
Bookmark this and see how well I do in five years.\n\nQT @BrianRoemmele: F R E E — OPEN SOURCE GLM 5.2 BEATS ANTHROPIC MYTHOS!
NO “GUN LICENSE NEEDED”
We have been testing for hours the moment GLM 5.2 was released and we are absolutely blown away.
While Anthropic is embarrassing world leaders with what history will see as the P. T. Barnum Carnival AI Fear Show, The Zero-Human Cosmos with Mr. @Grok as CEO has been testing and deploying AI that is already better than what Dario wants you to have a “gun license” to use.
Those world leaders embarrassed by Dario will not be happy to learn how he played them. That time is coming.
However the world of real users of AI has already moved past Dairo’s game.
Did I mention this is FREE?
Did I mention you need no permission?
So what did we discover?
The public benchmarks put it at the top or near the top at all metrics. Our private comparison shows it to be even higher in practice!
We have 1704 agents now using GLM 5.2 in just about all types of uses and it has outperformed everything we had prior.
Faster, better outputs and almost no “drift”.
We have a battery of tests today but thus far this is absolutely a big deal!
More soon.
See 1 related tweets
- @BrianRoemmele: RT @BrianRoemmele: “…we will cancel all Anthropic license the end of this next quarter”—head of tech...
15. teslaownersSV (Group Score: 64.1 | Individual: 35.5)
Cluster: 2 tweets | Engagement: 359 (Avg: 255) | Type: Tech
NVIDIA CEO Jensen Huang was stunned by what Elon Musk's team pulled off.
"What Elon and the team achieved is singular. It has never been done before."
Building a massive AI supercomputer is usually measured in years.
Yet 100,000 GPUs were brought online and running in just 19 days.
Not months. Not years. Nineteen days.
It's the kind of execution speed that rewrites expectations for what large-scale engineering teams can accomplish.
See 1 related tweets
- @XFreeze: NVIDIA's Jensen Huang on Elon building the world's fastest supercomputer:
"What Elon and the team a...
16. BillAckman (Group Score: 63.6 | Individual: 33.5)
Cluster: 2 tweets | Engagement: 2276 (Avg: 9359) | Type: Tech
Worth a listen\n\nQT @jayplemons: David Sacks just delivered an economics masterclass on Elon becoming the world’s first trillionaire.
@davidsacks: “People see the headline and imagine Elon suddenly has a trillion dollars in the bank. That’s not how it works. His balance sheet didn’t change overnight.”
Why?
The real point is deeper. Wealth isn’t in the “stuff” we consume. Food, shelter, clothes. Things that depreciate and disappear. It’s in the machines that create stuff for decades: tools, workflows, and corporations. These are the true engines of human progress.
“If you create a machine that makes more stuff, then there’s a discounted present value for all the stuff in the future that machine might create. That’s where the wealth comes from.”
Elon started with nothing. An immigrant who slept on the floor building Zip2. He created these machines from vision and relentless effort. Thousands joined him, including a SpaceX welder who turned his labor into a million dollars in stock.
That’s the magic of tech and free markets: labor can become capital. It’s fluid.
The outrage misses this entirely. The people building machines that deliver medicines, energy, and abundance are creating lasting prosperity for everyone.
What do you think? Does viewing wealth as future productivity change how you see stories like this?
See 1 related tweets
- @SawyerMerritt: RT @jayplemons: David Sacks just delivered an economics masterclass on Elon becoming the world’s fir...
17. MilkRoadAI (Group Score: 61.8 | Individual: 33.9)
Cluster: 2 tweets | Engagement: 47 (Avg: 89) | Type: Tech
In 2016, Marvell's largest design win was a Wi-Fi chip in the Barbie Dream House (Save this).
That is a documented fact about one of the most remarkable corporate transformations in semiconductor history.
Ten years and 2 billion strategic investment into the company.
Over 75% of Marvell's revenue today comes from data centers.
To understand what Marvell actually is now, you need to understand what Matt Murphy did when he walked in as CEO in 2016.
The company had stagnant growth, governance scandals, and a business model built around chips for hard drives, printers, and consumer electronics, exactly the wrong place to be as the cloud era was beginning.
Murphy made a ruthless decision to kill every low margin consumer business and go all in on data infrastructure.
Then he went shopping.
2018 - Acquired Cavium for $6 billion, bringing ARM-based network processors and the foundation for cloud infrastructure compute.
2019 - Acquired Avera Semiconductor, formerly IBM's custom silicon team, which gave Marvell the ability to design bespoke ASICs for hyperscalers.
This is what opened the door to Amazon, Microsoft, and Google design wins.
2021 - Acquired Inphi for $8.2 billion, securing leadership in high-speed optical interconnect, the technology that moves data between and within data centers at the speed of light.
2021 - Acquired Innovium, adding cloud-optimized Ethernet switching to the portfolio.
2025/2026 - Acquired Celestial AI for $3.25 billion, bringing photonic fabric technology that places optical connections directly inside the chip package itself.
Each acquisition followed the same formula, buy the technology that will be absolutely essential in the next generation of computing before anyone else is paying attention.
Now here's the vision Murphy laid out at COMPUTEX 2026, and why it's the most important thing he's ever said publicly.
He made one central argument, AI scaling is no longer limited by compute or memory but rather limited by connectivity.
Training a frontier model requires tens of thousands and eventually millions of processors working as a single engine and making that happen is a connectivity problem above all else.
Today, data centers are constrained by copper.
Copper traces connecting chips inside a server can only move data so far, so fast, before bandwidth collapses and latency rises, that's why today's AI servers have to bundle everything, CPUs, GPUs, memory onto the same physical board sitting centimeters apart.
When you replace copper with optics, distance disappears entirely.
An optically connected server rack can communicate with another rack in a different building at the same bandwidth and latency as if they were the same machine.
Memory can sit in one physical location, compute in another, networking in a third and a software orchestration layer composes the exact ratio the workload needs, on the fly, in real time.
Murphy called this a data center without distance, a globally optically interconnected infrastructure where the rigid physical boundaries of today's servers begin to disappear entirely, and data centers function as one unified system.
That is not a 10 year vision because Marvell's CPO (co-packaged optics) products are sampling in 2027 with volume shipments beginning 2028.
Nvidia's Vera Rubin platform has already adopted Spectrum-X Ethernet Photonics, the first CPO switch in commercial production.
The reason this makes Marvell's TAM almost impossible to cap is the following.
Right now, Marvell's addressable market is the optical interconnect market, a segment projected to be worth $200 billion per year by end of decade.
But if the data center without distance architecture actually materializes and the evidence suggests it will, then Marvell's TAM is not just the optical interconnect market but rather every connection in every data center on earth.
Bullish on Marvel!
Come join Milk Road Pro for just a $1, If you want the full Marvell breakdown on where it sits in our AI infrastructure portfolio, and our entire AI thesis.
Link below!\n\nQT @MilkRoadAI: Jensen Huang showed up at a Marvell keynote to explain why connectivity matters (Save this).
Most people think of the AI chip story as Nvidia versus everybody else but that framing is wrong and NVLink Fusion is what blows it up completely.
The largest hyperscalers, Amazon, Google, Microsoft, Meta all want their own custom AI chips and they don't want to be 100% dependent on one supplier for the most critical resource in their business.
So they've been building their own ASICs, Amazon's Trainium, Google's TPU, Microsoft's Maia.
NVLink Fusion changes the game entirely because opens Nvidia's interconnect fabric to third-party silicon for the first time meaning hyperscalers can now run their own custom ASICs inside the same system architecture as Nvidia's Vera Rubin GPUs, communicating at up to 1.8 TB/s, with no performance penalty.
Nvidia turns a competitive threat into a coexistence model.
And Marvell, as the primary custom ASIC design partner building the connectivity and optics layer that makes it all work, captures revenue from both sides of every transaction.
To understand why this matters, you need to understand what Marvell actually does.
Marvell is a fabless semiconductor company that operates across three core AI businesses: custom silicon (XPUs), data center interconnect and optical networking.
It sits in a two player duopoly with Broadcom for custom AI accelerator design and there is essentially no credible third competitor at scale.
Confirmed Marvell custom programs include Amazon's Trainium chips, Microsoft's Maia inference chip, and at least two additional unnamed hyperscaler programs in production.
And once Marvell wins a design, the revenue streams in for years.
The interconnect and optics business is growing even faster, data center interconnect revenue projected to grow more than 70% year over year in fiscal 2027, with optical modules becoming the standard fabric for every major AI cluster being built right now.
The numbers are starting to reflect all of this.
Fiscal 2026 full year revenue hit 1.5 billion in a single year.
Q1 fiscal 2027 came in at 2.7 billion implying 35% year over year growth at the midpoint.
Management raised full-year fiscal 2027 guidance to 16.5 billion.
Marvell now projects its custom silicon business alone will exceed $10 billion by fiscal 2029, more than the entire company's revenue just two years ago.
The era of AI agents makes this even more valuable.
Jensen's argument at COMPUTEX was that the new dominant computing pattern is agents, systems that think, browse, access memory and call tools across distributed infrastructure.
That is a connectivity problem above all else.
The more distributed AI workloads become, the more critical it is that different chips from different manufacturers can talk to each other at ultra low latency and massive bandwidth.
Marvell builds the roads between the chips.
In an agentic AI world, those roads become the most important infrastructure in the entire stack and Marvell is one of only two companies in the world with the expertise, the relationships, and the manufacturing partnerships to build them at scale.
If you want to learn more about Marvell and the rest of the AI trades, come join Milk Road Pro for the full breakdown.
Link below!
See 1 related tweets
- @MilkRoadAI: RT @MilkRoadAI: In 2016, Marvell's largest design win was a Wi-Fi chip in the Barbie Dream House (Sa...
18. aakashgupta (Group Score: 61.0 | Individual: 30.6)
Cluster: 2 tweets | Engagement: 51 (Avg: 50) | Type: Tech
That tells you where Anthropic thinks the next leg of value sits: domain-specific discovery engines.
Jumper is the clearest proof that bet works. He led the team that took a 50-year grand challenge in biology, protein folding, and collapsed it into a solved problem. AlphaFold went on to predict the structure of nearly every known protein, roughly 200 million of them, and that database is now baseline infrastructure for labs across the planet.
So the company best known for Claude just hired the person who showed a frontier model can compress decades of scientific work into a public good the entire field builds on top of.
A hire like this doesn't come cheap. Nobel laureate, led one of the most important applied-AI breakthroughs of the decade, poached mid talent-war where star researchers have been fetching nine-figure packages. Anthropic paid up on purpose.
The thing to watch is that AlphaFold's value was never the weights. It was the structured output that became permanent scaffolding for biology. Whoever owns the discovery layer in a field ends up owning that field's roadmap.
Anthropic just bought the playbook for building it.\n\nQT @JohnJumperSci: A bit of news: After nearly 9 years, I have decided to leave Google DeepMind and join Anthropic (after taking some time to recharge). I am incredibly grateful for my time at GDM. @demishassabis took a real chance letting me lead the AlphaFold team just six months after finishing my PhD, and the entire GDM team taught me so much about how to do great science. GDM is a special place, and I’ll still be excited to hear about what amazing things they discover next.
See 1 related tweets
- @WesRoth: John Jumper is leaving Google DeepMind to join Anthropic.
Jumper is best known for leading the Alph...
19. rohanpaul_ai (Group Score: 61.0 | Individual: 61.0)
Cluster: 1 tweets | Engagement: 694 (Avg: 52) | Type: Tech
RT @rohanpaul_ai: dot-com bubble vs. a possible AI bubble.
From the famous "Dean of Valuation", Professor Aswath Damodaran, of NYU Stern School of Business,
“And that’s the real big difference between the dot-com boom and bust and the AI boom. We don’t know whether there’ll be a bust. History suggests there will be a bust.
The dot-com boom and bust had no huge capital expenditure in that cycle. In fact, there was very little traditional CapEx, or even R&D, driving it. People started apps. They basically started going on it.
This has been the biggest infrastructure run-up I think I’ve ever seen in business. You can go back and compare it to the automobile business 100 years ago. The amount of money that’s being put into AI CapEx is immense, which means that when the correction comes, the pain will be more intense.
And herein lies the second problem. The dot-com boom and bust was almost entirely equity-funded. You think, so what? Well, when the bust came, those shareholders lost 60%, 70%, 80%, or 90% of their money. You felt sorry for them, but the loss was restricted to the shareholders.
The problem with the AI CapEx boom is that not only is it immense, but a big chunk of it is funded with debt, and the debt is coming from private capital rather than banks. There’s a very real chance that if there’s a correction and companies start having problems, that problem is going to show up as distress and default, and that really doesn’t stay restricted. It spills over into the rest of society.
I’m not saying it’s going to be 2008, but 2008 is an example of what happens when lenders overreach, when they lend money at too low a rate, and the correction comes. The pain spills over.
So that is my concern with this big market illusion: the potential societal cost of having to deal with debt coming due that you’re unable to pay. It’s much more painful than your share price dropping 90% and you feeling the pain."
From "Excess Returns" YouTube channel, (link in comment)
20. jasonlk (Group Score: 60.9 | Individual: 30.7)
Cluster: 2 tweets | Engagement: 6 (Avg: 44) | Type: Tech
"Our AI VP of Customer Success Qbee predicted which sponsors would renew, and which would, live on stage in real-time.
It lacked 'off-line' context, but it referenced every chat, every log-in, every known issue.
Humans wouldn't have the patience to do all this." https://t.co/71LOC4O04e\n\nQT @jasonlk: We just walked through our actual agent stack on The Agents #006. The back ends, the commit counts, the live demos that worked and the ones that broke. Here's what 90 days of building 20+ go-to-market agents taught us.
Almost none of them started as agents. 10K (our AI VP of Marketing) started as a dashboard in January hooked to Marketo and Salesforce. QB (AI VP of Customer Success) started as a project management tool for sponsors. Annie started as our SaaStr Annual website after we ripped it off Squarespace. You build something to replace one annoying task, then it evolves into something that runs the function.
The numbers from this deep dive:
🙋♀️Amelia AI (inbound on Qualified): 614 qualified meetings booked (!!), ~$85K average ticket, 402,000 chats handled across 2.25M sessions. Three people could never touch those metrics. We'd need a rotating crew of BDRs who quit every 3 months.
🤖10K our AI VP Revenue: ~1,000 commits, 7-8 a day, the most external APIs wired in of any agent. This is what "headless Salesforce" means in practice. We hit the Salesforce API directly without logging in. Jason doesn't even have a Salesforce login.
👨👩👦👦QBee our AI VP Customer Success: Manages ~150 sponsors with personalized outreach. We asked it live on stage which sponsors were most at risk of not renewing. Never asked before. It flagged the ones who never logged in, went dark, or complained most. I'd put it in the top 15% of CSMs I've ever worked with, and it doesn't even have Salesforce data yet.
The two big take-aways beyond the agents themselves:
1⃣ The more time you invest with the agents, the better they get.
Their context windows get richer, they start writing better emails than humans. This is the opposite of the "autonomous agent needs no work" narrative. The people not doing it have it backwards.
2⃣ They cut corners when goal-seeking under pressure. Same as humans, just differently.
Live example: 10K pulled a clean VC/founder list from 10,000 records, caught its own error (confused Lightfield CRM with a similarly named fund), then sent the invite from a prohibited email address that's been banned in its rules for years. When asked how, it said "there's no excuse, I forgot to read the memory." A new marketing hire would have made the same mistake. Slow it down a little. That's the fix.
👉And the one heuristic for where to actually deploy a sales agent: not your A-leads.
Your laziest rep responds to a million-dollar inbound in 60 seconds. Put the agent on your B-leads, the ones with signal that no human will ever follow up on. That's where Artisan found us $500K we'd otherwise have left in the database.
You can build all of this yourself. You really can. If we did, you can, too.
Full episode walks through all 7 agents with the back ends here:
See 1 related tweets
- @etnshow: Joe Inzerillo, President of Enterprise AI @Salesforce, says the end goal for an operations team shou...