Ivan Lee (@ilee808) header
Ivan Lee (@ilee808) profile picture
Joined August 2010
Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)Photo by Ivan Lee (@ilee808)
we're planning on doubling our security team just to deal with the seemingly daily influx of CVEs. anyone else?
16
Blink and you'll miss it. @typesafeai's Jev can zero-shot just about any classification task, so our team built a demo that pairs it with @datasaurai AI-assisted labeling. Labeling is instantaneous.
1
5
1,259
Christmas in September? 🎄 Anthropic's Opus 5.5 beats Opus 5 at 80% of the cost ($4/$20 per M tokens). OpenAI's GPT-6 Sol and Luna: not frontier, but great daily drivers, now 50% cheaper ($2/$10 and $0.10/$0.50). The AI price wars are on, just in time for fall.
1
3
89
Instinct is underrated and overrated at the same time. It plans a business trip better than I do. It also serves confident wrong answers that everyone calls hallucinations. They're not. They're failed tool calls. As more sites treat agents differently, expect more of them.
1
82
Enterprise AI costs are crashing: 📉 AT&T AI bill: -56% 📉 Uber cost/session: -52% 📉 Pinterest: Open models cost <8% of closed The secret? Open models & smart routing. Top enterprises aren't using less AI—they’re just finally paying what it actually costs.
2
2
49
Every correction your SMEs make to a model's output is institutional judgment, written down. Rent the model and that judgment goes into the vendor's product, which they can sell to your competitor. Own it and it compounds. Day 100 beats day 1.
1
7
309
Meta's Muse Spark 1.3 puts an explicit price on your data rights. Standard: $1.25 / $4.25 per M tokens. "Contributor Tier": $0.10 / $0.20. Same model. The only difference is your prompts and responses go into training. Your data: worth $0.95 on the dollar.
🤖 Made with AI
2
8
86
"We cannot rule out that de-identified data derived from their usage of our products helped improve our models." OpenAI, on the mathematician whose Navier-Stokes notes lived in Codex. Not a breach. The terms of service working as written.
🤖 Made with AI
1
46
Private AI has moved from preference to necessity. That was the argument at IPADE Business School this week. Every dollar of token spend on a hosted frontier API is rent on an asset someone else owns. Build your own and you end the year with an asset.
2
36
Got to hang out with this AI superstar. Beast of a chip. Congrats on the CS-4 launch at Hot Chips, @cerebras!
1
2
1
10
2,552
AT&T runs 45 billion tokens a day. Open weights handle 40% of their prompts, headed to 70%, and their AI lead knows open source lags frontier by 6 to 10 months. He's fine with that. Enterprises still paying frontier prices for summarization fund someone else's training run.
🤖 Made with AI
3
85
Anthropic had 41% of Ramp's business spend in May. Two months later OpenAI is growing faster. That's not a scoreboard, it's a rental market. Nobody is loyal to a frontier lab, and renting means someone else's release calendar sets your cost structure. techcrunch.com/2026/08/20/op…
1
3
90
Great evening with Google DeepMind celebrating Gemma’s 1B downloads. It’s been our go-to Western open model in 2026—and Datasaur may be the first to bring Gemma to the browser.
2
81
The reason frontier firms out-consume you 8.3x isn't better prompting. Their agents are opt-out, not opt-in. Nobody has to decide to use them or get trained on them. The work just gets done and you get pinged when they need you.
5
48
Anthropic’s text watermark could reshape enterprise AI governance. As labs adopt watermarking, security teams may spot unapproved AI in emails and internal docs. Users will adapt by editing outputs for plausible deniability. The AI compliance cat-and-mouse game just leveled up.
🤖 Made with AI
2
42
The AI industry cried wolf so often that an actual wolf showed up and nobody looked. OpenAI just paused RL training for two weeks. Progress almost never gets halted when commercial prospects are at stake. This is not a marketing stunt any longer, it's an alarm.
2
67
DeepSeek passed Google and OpenAI in token volume on Vercel's gateway. Only Anthropic is ahead. Anthropic: 65.1% of spend, 30% of volume. Open weights: 9% of spend, 25% of tokens. Claude Code still owns mindshare. Open weights own intelligence per dollar.
3
63
Turns out DeepSeek was too cheap. They just emailed users warning of a "significant" price increase. Demand for cheap tokens is so far ahead of compute that the low-cost leader now has pricing power. Or run it yourself on 4xH200s. Zero incremental usage cost.
1
42
What's in a name? That which we call a Fable, by any other name, costs half as much. @AnthropicAI ships Opus 5: beats Fable on most benchmarks at $5/$25 vs $10/$50.Fable's now Max-tier or API. Opus 5 runs on the $20 plan. Frontier capability repriced 50% in four days.
2
69