CEO @with_cortex | Building systems for AI to understand us beyond prompts
Palo Alto, California
Joined November 2019
- Tweets1.7K
- Following136
- Followers2.3K
- Likes1.8K
Full-Duplex-Bench-v3 tests voice agents on multi-step tool use with real human speech full of ums and mid-sentence corrections. GPT-Realtime is first at 60% Pass@1, then Gemini Live 3.1 at 54%. Even the leader fails on over 40% of self-correction scenarios (Full-Duplex-Bench-v3)
Hot take: a voice agent should keep a running list of what you already corrected mid-call, so it stops booking the first city you said
Got a good tip on a caregiving forum this week. If you need free Medicare help, google "SHIP counselor in my area", most senior centers have someone trained and it's easy to book. KFF says in 2021 a one-on-one SHIP session averaged 33 min vs 9.5 for a call to 1-800-MEDICARE
Pew asked Americans if AI will be good or bad for them personally over the next 20 years. 29% of adults 65+ said they're not sure, more than said negative (28%) or positive (17%). For 18 to 29 year olds it was 12% (Pew Research Center)
Feedback Bench ranks 12 personal agents from what their users post on Reddit, X, YouTube and the app stores, with 50 as the average agent. Muse from Meta is first at 73.4, then Hermes Agent at 62.5 and Grok Bot at 57.9. ChatGPT Dots is last at 38.5 (Feedback Bench by Enterpret)
Hot take: a personal agent should send you a weekly summary of what it did for you, like a bank statement
Someone on a caregiving forum told me they found their fiduciary by googling "fiduciary" plus their town. They'd been nervous the firm wouldn't want them, then realized it was courting them as a client. 56% of caregivers of adults 50+ help manage finances (AARP/NAC)
In AARP's latest tech survey 43% of adults in their 50s had used generative AI, and 14% of adults 80+. Work seems to be a big part of it, 42% of employed adults 50+ had used it vs 21% of those who aren't employed
Artificial Analysis puts voice models through replica customer service calls (airline, retail, telecom) and scores how many get resolved. Phonic 1 (High) now has the top spot at 79.9%, then GPT-Live-1 at 74.5% and Grok Voice 3.0 at 71.6%. GPT-4o Realtime from Dec 2024 is at 27.9%
Hot take: before an agent pays for anything it should read the order back to you, like every drive thru does
labs spent all year fighting distillation through their APIs. now people just have Astra edit Qwen's weights directly and call it programming 😂 teapawt.net/rikiyatakehi/status/21…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Was reading a Medicare thread last week where someone's eye drops were $650 list, $365 through their plan and $79 through the manufacturer's program. They found out about it from a woman at pickleball. KFF says 65% of people on Medicare didn't compare plans in the last open enrollment
U-Michigan asked adults 50+ who've used AI for health info whether it beats asking a person. 47% said the person is better, 10% said AI. Makes sense to me, with health stuff you usually want someone who can ask you follow-up questions
Artificial Analysis now tests whether voice models actually stop when you talk over them. Most new ones are in the high 90s, Eleven Realtime v1 Max 99% and Gemini 3.8 Live 97.5%. Gemini 3.1 Flash Live Preview handled 4%
Hot take: voice agents should just say "hold on, let me check" like a normal person when a tool call is slow
