Sabitlenmiş Tweet
DARIO ESSAY | AI POLICY | OPENAI-ANTHROPIC PRICE WAR x.com/i/broadcasts/1…
English
MTS
1.6K posts
SITUATION EXPLAINED: Why is effective compute growing so fast?
• Hardware manufacturing improvements (Moore's Law): 1.5x per year, for six decades straight
• Investment in compute hardware: 2.5x per year, bigger than Moore's Law itself
• Combined, these two have driven ~4x per year growth in training compute for the last decade
• Algorithmic efficiency improvements add another 3x to 6x per year on top of that
@theojaffee: "Four, sixteen, sixty-four, two fifty-six, and it just keeps going and going."
MTS@MTSlive
SITUATION EXPLAINED: Effective compute for AI training is growing at roughly 10x per year. • "Effective compute" combines three things: more hardware, more spending on that hardware, and algorithms getting more efficient • Year 1: 1x. Year 2: 10x. Year 3: 100x. Year 4: 1,000x. Year 5: 10,000x • That's a 10,000x improvement in 5 years • A million x improvement in 7 years • A billion x improvement in 10 years • Epoch AI says 10x per year is on the low end of estimates, the actual rate may be higher and accelerating @theojaffee: "Do you understand how insane it is for anything to grow an order of magnitude per year?"
English
SITUATION EXPLAINED: Effective compute for AI training is growing at roughly 10x per year.
• "Effective compute" combines three things: more hardware, more spending on that hardware, and algorithms getting more efficient
• Year 1: 1x. Year 2: 10x. Year 3: 100x. Year 4: 1,000x. Year 5: 10,000x
• That's a 10,000x improvement in 5 years
• A million x improvement in 7 years
• A billion x improvement in 10 years
• Epoch AI says 10x per year is on the low end of estimates, the actual rate may be higher and accelerating
@theojaffee: "Do you understand how insane it is for anything to grow an order of magnitude per year?"
English
Every consumer product has to rethink its interface.
@andyfang, co-founder of DoorDash:
"For the first time ever, agents are more web traffic than humans, whether we like it or not, we need to figure out a way to change our interfaces."
"I used to think exactly the same way Brian Chesky did, I don't want a conversational experience in the product for consumers."
"But people's behaviors are changing, Google has publicly stated that search query lengths have only gone up over the past couple years. People are trying to speak to it like a human versus translating what's in their head into keywords."
"We've grown up doing that keyword translation for decades, we just assumed that's the most natural way to interact with computers. But now that these AIs are here, people are just trying to speak to it like a human."
@julesrosenberg
Andy Fang@andyfang
Today we’re launching Ask DoorDash — a new conversational way to search the app in your own words through chat, voice, a recipe link or photo. Ask DoorDash can build you a grocery cart ~5x faster than doing it manually. It takes a single prompt to complete your cart in under 2 minutes. In early testing, nearly half of all restaurant orders made with Ask DoorDash were from a place the customer had never ordered from before, and grocery baskets built with Ask DoorDash were over 35% larger than those without. [1/3]
English
MINI SITUATION RECAP:
OpenAI plans to undercut Anthropic’s pricing.
In light of massive Anthropic revenue growth and enterprises’ desire to cut token spending, OpenAI is considering lowering prices across all its models. If OpenAI significantly cuts prices, it could eat into both Anthropic’s enterprise share and both labs’ profit margins.
English
SITUATION EXPLAINED: How does the jump in capabilities for Fable 5 compare to previous model releases? @ml_angelopoulos from @arena explains.
"Yes. I’d say we have seen precedents for this level of improvement.
Not yet on Agent Arena, which is a new benchmark, so there hasn’t been enough time to build up the historical data needed to understand how often jumps like this happen. But certainly, when GPT-4o came out, it represented a huge leap in performance. Earlier this year, the high Claude 4.X models were also massive leaps forward."
"Another major improvement, perhaps the largest jump in model performance we’ve seen, has been in multimodal capabilities with GPT Image 2. When that model was released, it caused a tectonic shift in the leaderboard, and it still maintains a Bradley-Terry coefficient gap of more than 100 points over the next-best model."
English
SITUATION ANALYSIS: Prometheus Announced
For the last year or two, Jeff Bezos has been working on a stealth AI startup.
Today, he announced the next stage of neolab Prometheus, with a $12 billion fundraise at a $41 billion valuation. Backers include Bezos himself along with J.P. Morgan, BlackRock, Goldman Sachs, DST Global, and Arch Venture Partners. Bezos and scientist / investor / entrepreneur Vik Bajaj will be co-CEOs.
This is Bezos’ first time in a CEO role since stepping down from Amazon in 2021. Bezos is busy: he’s the executive chairman of Amazon, owner of Blue Origin and the Washington Post, and an active angel investor.
Before Prometheus, Bajaj was the co-founder and CEO of Foresite Labs, a startup incubator. He ran Foresite’s investment arm, Foresite Capital, and was CEO of one of Foresite’s companies, Xaira Therapeutics. Before that, he was the chief scientific officer of cancer detection startup Grail. Before that, he was a director at Google X, Google’s “moonshot factory” R&D organization, where he led Google Life Sciences, now known as Verily.
Prometheus’ mission is to build an “artificial general engineer” that will ingest a huge amount of data from the physical world to accelerate engineering and manufacturing in a range of fields. Once AGE has been achieved, Prometheus plans to raise tens of billions for a holding company that acquires existing companies or incubates their own from scratch, then enhances them with Prometheus’ AI — sort of like a Berkshire Hathaway for the 21st century.
Via @theojaffee
MTS@MTSlive
SITUATION BREWING: Prometheus, the industrial AI startup led by Jeff Bezos, is set to announce a $12B Series B at a $41B valuation, per Axios. The company is building what it calls an “artificial general engineer.”
English
SITUATION EXPLAINED: @ml_angelopoulos, CEO of @arena joined us to explain how Claude Fable 5 is performing.
"What this means is that, compared with the average model, Claude Fable 5 is getting about 11% more positive signals from real agentic workflows. In the middle, models like GLM and Gemini 3.5 Flash are roughly at baseline, and then performance gets worse as you move down the leaderboard."
"The signals on this leaderboard come from a few different sources. The first is task success rate: explicit signals where users say they were able to complete the task using a given model.
Fable has an 18% task success rate. By comparison, the next-best competitor, Claude Opus 4.8 Thinking, is at 9%. So Fable almost doubles the effect of the next-best model. That’s a massive improvement in people’s ability to get the job done."
English
SITUATION EXPLAINED: Broadcom stock is down nearly 20% from highs this month. @StevenGlinert, CEO of Sphere Semi, explains what Broadcom does, why they matter, and what might have spooked investors.
"Broadcom is a funny company because they do a bit of everything."
"From a data center perspective, Broadcom’s role is close to being a chip architect, a high-speed I/O supplier, a networking supplier, and a manufacturing translation layer all at once."
"People in financial markets looked at Broadcom’s growth and expected it to be stronger. So there are really two layers to the reaction.
The first is that Broadcom is growing, but not fast enough relative to expectations."
"The second layer is that Broadcom’s growth can also be read as bearish for certain parts of the compute market, especially Nvidia.
In some ways, buying Broadcom is a bet against Nvidia. It suggests that customers may want to stop paying the “GPU tax” and instead build or use alternative chips, which means paying the “Broadcom tax” instead."
English
SITUATION EXPLAINED: Joshua Levine (.@JoshuaTLevine) of @JoinFAI argues that a quantum buildout with onshore supply chains will drive jobs and growth in the US while also addressing the risks of relying on strategic adversaries.
"The key issue for quantum is the supply chain behind the machines themselves. That includes things like photonics components, rare earth processing, lithium niobate, helium-3, quantum-grade wafers, and advanced packaging.
Some of this should sound familiar, because it overlaps with what we have seen in next-generation chips and chipsets for GPUs, driverless cars, and other modern technologies."
"The difference is the purity level required for quantum systems. Traditional semiconductor manufacturing already requires very high purity, but some quantum systems need even more. We are talking about three to four nines of purity.
That means the total scale of supply may be smaller than in semiconductor manufacturing, but the quality has to be extremely precise."
"Because demand has historically been limited, there are not many suppliers. And many of the existing suppliers are based in adversarial nations, particularly China."
Joshua Levine@JoshuaTLevine
Excited to share that @PrinehaN and I had the most-read Cogs of War piece on @WarOnTheRocks in 2025. The piece, America’s Quantum Manufacturing Moment, kicked off our four-part series on policies and strategies to promote US Quantum Leadership. A quick 🧵
English