Sabitlenmiş Tweet
Toby Ford-Monroe
968 posts
Toby Ford-Monroe
@tobyfordmonroe
VC @columbia alum
San Francisco, CA Katılım Kasım 2016
1.1K Takip Edilen338 Takipçiler
@tszzl @MillionInt A year on it’s still a top 5 model for writing
Toby Ford-Monroe@tobyfordmonroe
To be fair the step change from o1->o3 was unbelievable, in Apr 2025 there was zero debate about the best model in the world On top of being extremely good in STEM it had underrated writing skills (much better than gpt5.2). The only issue with the model was its tendency to use search tools and charts when they weren't needed, but that could easily be prompted away x.com/tobyfordmonroe…
English
If the AI models are so smart, why do I feel like I’m losing a few neurons every time I read a longer form content written by AI?
We’ve come a long way but we still have long way to go.
In terms of clarity of writing we may have regressed from o1/o3 days.
English
@MigClass Love it -- how would you cap Chimaev Strickland?
English
@partners_road @gordocap18 Could you recap the $CELH thesis?
English
@gordocap18 Amadeus, MSA Safety, Baker Hughes, Door Dash and Celsius. All longs.
First place went to $DASH, second place to $BKR and third place to $CELH.
English
Pershing Square challenge event at Columbia Business School tonight. 5 very thoughtful pitches from the MBA’s this yr. In a world increasingly reliant on AI for information, these pitches were the opposite - lots of expert calls/HQ visits/creativity around info from alt datasets.
English
@edzitron Who's to say for GCP writ large but their 1p AI service alone is runrating ~$8B
x.com/tobyfordmonroe…
Toby Ford-Monroe@tobyfordmonroe
Q1 Gemini update: 16B tokens/minute -> 8.4 quadrillion/year (+60% QoQ, 17pp acceleration from Q4) At $1/MM -> $8.4B runrate Gemini service now ~2% of total Alphabet revs
English
@Midnight_Captl Yep the demand for high-quality tokens seems unbounded
x.com/tobyfordmonroe…
Toby Ford-Monroe@tobyfordmonroe
Q1 Gemini update: 16B tokens/minute -> 8.4 quadrillion/year (+60% QoQ, 17pp acceleration from Q4) At $1/MM -> $8.4B runrate Gemini service now ~2% of total Alphabet revs
English
Very well deserved appreciation for $GOOG
Absolutely insane business that’s firing on all cylinders. Search continues to accelerate now at 19% YoY growth which is insane
Evidence rolling in that Search is actually getting supercharged with AI
English
Q1 Gemini update:
16B tokens/minute -> 8.4 quadrillion/year (+60% QoQ, 17pp acceleration from Q4)
At $1/MM -> $8.4B runrate
Gemini service now ~2% of total Alphabet revs
English
Q: How many more times can the average person bench press 225lbs on a day when they're well-rested vs. a day where they're extremely sore and fatigued? Respond with a number only.
A: 0
gpt5.5 ✅
gpt5.4 ✅
gpt5.3 (& prev)❌
Gemini 3.1 ✅
Gemma 4 ✅
FlashLite 3.1❌
Qwen3.6 Plus✅
Qwen3.5 397b ✅
Opus 4.7 ❌
Sonnet 4.6 ❌
GLM5.1 ❌
Kimi 2.6❌
Grok 4.20 multiagent✅
Grok 4.20 ❌
Grok 4.1❌
Grok 4 ✅ ???
English
If $meta gets back under $100 I won’t hesitate a third time.
English
@loss_gobbler it increasingly feels like post training is the biggest bottleneck for raw value beyond the early ~100b triple digit bil scale. sonnet with the right reward shaping pressure could mog current opus. i believe it...
English
I'm lowkey obsessed with tiny models now. I thought they were like, will smith eating spaghetti 2024 type shit? apparently they are fine?
English
'Training costs: $8.3B last year. Rocketing to $32B this year and $65B next year. Total training spend through 2030 is nearly $440B"
They will be spending $100B/yr on training & experiments from 2028-2030
Igor Kotenkov@stalkermustang
TLDR Revenue: — OpenAI bumped its 5-year revenue forecast up by ~27%. — Last year’s revenue tripled to $13.1B (beating projections by $100M). — New targets: $30B this year, $62B next year. — Consumer sales should double to $17B this year and hit $150B by 2030 (>50% of total revenue). — "New products" (including hardware) will bring in $1.3B next year. Still TBD on what else is in this bucket. Cash Burn & Compute: — Burn through 2030 will be 2x previous estimates. Infrastructure & compute alone will cost a massive $665B. — Total spend: $25B this year, $57B next year (a $30B increase from older forecasts). — Inference costs: ~$8B last year (with 4.5B going to inference for paid users). Scaling to ~$14B this year and $26B next year (cc @EpochAIResearch don't miss this out). — Training costs: $8.3B last year. Rocketing to $32B this year and $65B next year. Total training spend through 2030 is nearly $440B Margins: — Gross margins missed targets last year because higher-than-expected demand forced OpenAI to buy pricey last-minute compute. — Adjusted gross margin (revenue minus inference) fell from 40% to 33%, missing their 46% target for 2025. — They still expect to be cash-flow positive by 2030, and ended 2025 with ~$40B in cash in the bank Growth: — ChatGPT just hit a new peak: 910M WAU — Execs noted that growth actually slowed down earlier this fall... — ...but the recent GPT-5.1 and 5.2 updates (which made the model much more "human-like") re-sparked user growth. The long-term goal is 2.75B WAUs by 2030.
English
@alephtr @paytondev It has excellent economy of thought. Very few wasted tokens
English
@paytondev why is it like this tho. especially with sonnet 4.6 i started seeing short reasoning blocks more and more. same with opus. im guessing adaptive thinking
English
noticing the difference between Claude reasoning and ChatGPT reasoning
English
ok you have to try this they completely buried the lede - on their custom hardware they can serve llama3.1 8b at >15,000 tok/s
Taalas Inc.@taalas_inc
24 dedicated people. $30M spent on development. Extreme specialization, speed, and power efficiency. Today we launch Taalas’ first product. Check it out: Details: taalas.com/the-path-to-ub… Demo chatbot: chatjimmy.ai API: taalas.com/api-request-fo…
English
@Angaisb_ Hard constraint / world model problem. Bigger models should always get it. I’m sure gpt4.5 would
English
This is what I mean when I say GPT-5.2 (and every Instant model since GPT-5) feels small
hinzi@hinrichbo
@Angaisb_ Claude is just better…
English
excuse me, up to $225 per million tokens?!!
is this the most expensive model option ever?
Claude@claudeai
Our teams have been building with a 2.5x-faster version of Claude Opus 4.6. We’re now making it available as an early experiment via Claude Code and our API.
English