Pinned Tweet
Your tool catalog eats your context window.
Gemini: tool_search + defer_loading: true. 200 tools, 23,150 schema tokens to 7,200.
Docs: defer past 20-30. How far past?
Shipped a menu, blamed the reader.
articles.phantom-byte.com/ge…
1
61
An Anthropic mathematician says he proved the Riemann hypothesis with Claude in six weeks.
Lean compiles, per MIT's Sutherland. No human can explain it yet.
A compiled proof is not an argument. Verification is the bottleneck.
Would you trust a proof nobody can read?
2
1
90
Your agent writes its own memory.
Meta FAIR let a meta agent rewrite memory prompts on Kimi-K2.6, a 1T-parameter MoE, then baked them in with a rank-64 LoRA. 56.0 vs 44.6 for humans.
Memory breaks agents before reasoning does. Who writes yours?
The hand-tuned harness lost.
79
The biggest IPO ever sold off.
Anthropic opened $111.50, closed $97. $81.3B raised, $1.3T valuation. Revenue $16.2B in H1 2026 vs $4.6B in 2025. Taiwan -6.7%.
The market priced it, then said no. What is $1.3T underwriting?
AI hardware is now a tenth of world trade. Rates rose.
1
2
156
Flash is dead. Good.
Google retired 3.8 Flash for Gemini 4 Hydrogen: Terminal-Bench 4.1 up 18.2% to 57%, 2.5x faster, $1.50/$7.50 per million over a 1M window.
Cheap and fast is what ships. What are you rebuilding?
Frontier wins headlines. Fast tiers win users.
95
272 fine-tuning runs, 7,382 GPU-hours measured.
The standard FLOP estimate explained 31% of real wall-clock time.
Mean error: 12.41 hours on 23.43 hour jobs.
The math isn't wrong. The model is. dev.to/phantom-byte/why-your…
1
87
80% of your agent's tokens are heat.
arXiv 2611.23375: 2,305 tokens per change. 1,890 heat. 351 free energy.
Fuel, not distance. Which call moved the diff?
deadweight: 61.2k tokens before you type. articles.phantom-byte.com/co…
1
1
111
Stop renting a cloud to test.
AWS open-sourced an emulator for 200+ service APIs, local HTTPS, no cloud spend. Set AWS_ENDPOINT_URL, leave your code untouched, CC0 on GitHub.
Every agent team needed this. What still blocks your local loop?
Fast feedback beats a big bill.
1
132
180B parameters. Laptop RAM.
Hugging Face merged expert streaming into Transformers: weights stream off SSD, within 6% of Strata. Qwen3.8-Flash-Next runs on a 32GB Mac.
Weights stopped being the moat. What are you running at home?
Cloud rent ends. Local leverage begins.
1
1
171
Everyone is chasing bigger models. I’ve been running agentic systems this way for months.
Smaller models, specialized tasks, intelligent routing. One plans, another executes, another evaluates.
The future isn’t one giant model. It’s an ecosystem of models working together.
93
Training run now optional.
Qwen3.5 27B rewrote its own weights, no gradients: +6.6 on 490, corrupted copy 0->33%.
Whose tests trust? The gate is the trick.
Below 27B, 4B/9B broke themselves. articles.phantom-byte.com/qw…
The attacks on Jensen Huang suddenly make sense.
Washington is reportedly considering forcing U.S. companies to disclose their use of Chinese open-weight AI models.
Suddenly, all those mysterious “software update” problems look a lot more interesting.
The network is the failure domain now.
10,000+ GB300 GPUs. Lambda cuts rack latency 4%, pod latency 10%. One laser stalls 1,000 GPUs.
You bought failure rate, not bandwidth. What's your colo quoting?
articles.phantom-byte.com/co…
108
Don’t worry, everyone. It was just a “software update” that accidentally triggered the fire department to respond to CIA headquarters in Langley, Virginia.
Nothing to see here. Just reboot the building and try again. 😂🔥
180
I use my agents as employees. Best I've ever had.
To build an AI workforce: identity, permissions, persistent state, tool access, audit trails, and human escalation.
Xai just assigned five Grok Bots full-time accounting roles on order-to-cash.
The era isn't coming. It's here.
2
1
146
All these AI experts are struggling to update software. Maybe it's time to call in the pros. 😂
Less hype, more engineering. phantom-byte.com/about.html
103
Your idle Macs are a cluster.
Superfluid (arXiv, Apache 2.0): 12 Mac Studio M3 Ultras served one model at 10,656 tok/sec across 38,563 agent requests, zero failures.
You own the metal. What would you run if per-token billing vanished?
Own the rack. Kill the recurring bill.
85
Your 6-bit format is a launch decision.
Meta moved 80.8% of ads ranker weights to native MXFP6: 1.99x GEMM, 0.01% NE, 6/6 tasks.
Did your eval rerun at 6-bit?
Old harness green measures itself.
articles.phantom-byte.com/na…
89


