Karim Mattar (@MattarARK) header
Karim Mattar (@MattarARK) profile picture

AI Research Associate @ARKInvest | AI, Cloud & Semis | Research x Automation | https://nitter.cf/t.co/TvxuqE2drz | Disclosure: https://nitter.cf/t.co/7SQwpFZRTf.

Joined September 2024
Looks like @huggingface is about to move 1 exabyte a month. That's a lot of open models getting pulled down to run somewhere
we are this 🤏 close to sending 1 EB (1 EXABYTE) of data monthly over the internet We'll keep optimizing our platform to distribute intelligence globally, w/o gatekeepers.
180
Same score, same price Fable 5.2 from @AnthropicAI on low and Gemini 4 Hydrogen from @GoogleDeepMind on xhigh both land at 54, about $3.10 vs $3.16 a task. Drop Hydrogen to low and it's $0.29 a task
Replying to @ArtificialAnlys
xhigh nearly doubles output tokens compared with high for a 3 point gain in Intelligence, and Cost per Task falls from $3.16 at xhigh to $1.87 at high. For most workloads, high gives Argon-level intelligence at a fraction of the cost.
324
GPT 6.2 Astra does the task for under a third of GPT 6 Astra's cost just 75 days later, faster than the 95% a year inference cost decline in ARK's Big Ideas research
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…
304
He had @AnthropicAI Opus 5.5 audit 50,000 astrophysics papers and it found nearly 1000 major errors. Cost was about $200 of Claude Max plus 45 hours of his time. Peer review now has a super cheap second checker
In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included). Bottom line: AI can speed up quantitative science by >1000x with no further model improvement required. But I'm seeing no evidence of this being pushed within universities, only academics complaining about labs dropping math proofs. Among my fellow reformed academic high agency founders, however, several are speed running entire disciplines. This survey has implications for how science will be run in the very near future. Hint: Academic publishing is dead. It was already a very stupid way to record and promulgate knowledge, and had been so for decades, even before all that publicly funded knowledge was locked behind paywall cartels. Now AIs can generate thousands of novel, correct papers per day, it is no longer meaningful for humans to consider a research paper a "unit of science" or to even bother much with the 99.9% of science which, up to this point, was necessary drudgery. Time to discover something like General Relativity or Evolution every week. LFG!! Enjoy: science.caseyhandmer.com/phy…
1
361
Karim Mattar retweeted
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…
2,174
3,552
3,606
35,089
8,841,714
🤖 Made with AI
Jensen Huang says he would return as $NVDA CEO if the board asks in his first public statement since stepping down Thursday. He says he "owns my mistakes completely."
3
3
423
Last year was already a record for US power generation. The November EIA outlook puts 2026 at almost 4,390 billion kWh, up 2.7%, with commercial sales passing residential.
286
How do you grow fast and get cheaper at the same time? In my @AnthropicAI section this week: revenue per dollar of compute is up more than 4x in 18 months, helped by a chip stack spanning Nvidia, AMD, Google and Amazon, and Opus 5.5 cut cost per task about 40% vs Opus 5 Check it out!
5
2
13
6,902
15 links 6 OTPs 3 applications SSN and birth certificate just to get into my grok bot account
1
1
3
545
So @OpenAI says 1,800+ researchers reported nearly doubling their productivity with ChatGPT, and now it wants up to 200,000 scientists on frontier models for free We could see the cost of a discovery drops fast if this holds up!
Scientists reported nearly doubling their productivity with AI in our study of more than 1,800 participants in the ChatGPT for Academic Researchers program. Using privacy-preserving methods, we looked at how they use ChatGPT across their work—from developing ideas to analyzing data and writing code—and what it could mean for scientific discovery. openai.com/index/science-at-…
1
315
The amount of times I’ve heard an always on agent tell me “you’re right I forgot to do xyz even though we talked about doing it every day”
2
395
Karim Mattar retweeted
There's a million ways to cut this so don't fixate on one. Ex: A 20% token-price cut needs 25% more tokens sold to keep revenue flat: 1.25 × 0.80 = 1.00. The earlier 1 GW example held volume fixed. Allow tokens per GPU-hour to improve, and the range of outcomes changes. With prices down 20%, on the same fleet, GPU-hours and spending: >Same sold volume: operating profit falls from just under $4B to a small loss. >Throughput rises 25%, all extra output sells: revenue stays around $20B and profit just under $4B. >Throughput rises 25%, but sold volume grows only 10%: revenue falls to about $18B and profit to $1.3B. TLDR = Faster throughput helps when the additional output sells. If producing that volume requires more GPU-hours or new hardware, profit must also absorb those costs.
To illustrate how quickly it all can change based on the token price assumption, spun up a quick example. At an illustrative 1 GW AI company, a 10% cut in realized token price can more than halve operating profit, from just under $4B to below $2B. A 20% cut can wipe it out, even with ~$7B of EBITDA left. This is a simplistic exercise with clear limitations but can show the directional impact of how token price swings impact end profitability and why folks want to hedge or procure bilaterally under long-term contracts. The company owns the campus, GPUs and model business: about $50B of capital and 400,000 GPUs selling roughly 16 quadrillion billable tokens a year for about $20B of revenue. Cash costs are about $9B, leaving $11B of EBITDA. ~$7B of annual depreciation, assuming five-year IT and 30-year facility lives, doesn't change with what tokens sell for. So with spending fixed, price changes flow almost straight to operating profit. Profit hits zero at a ~20% price cut, or 25% if company costs fall in step with revenue. One thing folks miss too about power, electricity is a much smaller lever. The bill is <5% of total costs, and halving the power price adds about $360M a year against roughly $2B for a 10% token-price move. Paying up for power can still make a lot of economic sense. An extra $100/MWh costs about $700M a year. A finished cluster waiting for power has an EBITDA run rate of ~$1B a month, so one extra operating month covers about a year of that premium. Three more points of productive utilization +$0.7B a year with company spending fixed, which can support an ongoing premium.
1
10
44
6,179
Fable 5.2 from @AnthropicAI now clears 41.25% of real freelance projects on @scale_AI's Remote Labor Index. Second place is Opus 5.5 at 21.25%. A year ago the best agents sat around 2.5%.
5
5
498
All these always on agents (grok muse duet dots) need a better way of communicating and visualizing. Give me a dashboard or something 10 messages back to back from each one is just stressing me out
3
7
692
My 200$ GPT plan should never be running out on Terra high, let alone within the first 3 days of the week
2
1
1,215
So the companies leaning hardest into AI are the ones hiring faster. Hard to square that with all the job loss headlines
Cathie Wood says ARK expects a labor shortage, citing baby boomer retirements, unemployment dropping toward the 3% range, and Ramp data showing AI-intensive companies hiring faster. She says warnings of massive job losses helped set the stage for the AI backlash.
1
1
440
Starship flight 15 is live. Roughly 12 million pounds of propellant loading right now 🚀
Live webcast of Starship's 15th flight test starts now → nitter.cf/i/broadcasts/2brajgZlO…
1
4
660
Generate a 3D model then use any angle for the next video. All in one place now
3D is now supported in Runway. Generate detailed 3D models from text or images with Tripo H3.1, or upload your own. Choose any angle and use it as a reference for your next image or video. Export your model to use in other 3D tools. Try it now in Runway.
377