Jeff Barg
@jeffbarg
Followers
2K
Following
39K
Media
105
Statuses
3K
AI @ @clay_gtm • prev https://t.co/LmhUuxxUhq (YC W21) @amazon @pennmandt
New York
Joined April 2009
What does the future of agentic AI hold? Join us next Monday for a panel with @clay_gtm and @vercel. We'll cover topics like patterns for building agentic apps, multi-step workflows, and the evolving toolchain for devs. 📅 Sep 8, 6:30PM @ Clay HQ RSVP: https://t.co/OKRqwIxj1L
1
1
4
My only contribution to the evals discourse: Evals are only good insofar as you can actually look at failures and understand them. The actual eval pass rates are pretty meaningless in isolation.
2
0
10
If I were in charge of model naming for OpenAI, the few adjustments I have would be: gpt-4.1 → gpt-4a (4-agentic) gpt-5-high → gpt-coder
0
0
2
We’ve been working on a similar transition for the AI features at @clay_gtm. I’d characterize the shift as “tell, don’t show” and rely on model reasoning.
A lot of work goes into defining tool calls for agents. Moving the team from thinking about few-shots (gpt-4 era) to tool definitions (sonnet4/gpt-5 era) has been a company-wide initiative that has unleashed major step-change capabilities. Advice - organize your work around
1
0
9
"what gets measured gets managed" is effectively restating "what can be verified gets rewarded"
0
0
2
The whole push around reinforcement learning for making great coding models presupposes that reviewing code is easier than writing code. From experience, I don't think that's a given.
5
0
12
hey can everyone stop using Sonnet for 30 minutes? trying to do a demo and would love the extra bandwidth, thanks!
6
0
56
I was just thinking about how agents (even text-based ones) will probably look a lot more like voice agents in the future. Things that are important: - Interrupts - Partial messages and streaming input - Thinking fast and slow
The Realtime API is GA and with it we introduced a lot of new features and a new gpt-realtime model 🎉 📝 Better instruction following ⏳ Async tool calling 🖼️ Image input 🧰 MCP tool calling 📞 SIP support 🎓 A new comprehensive prompt guide to get the most out of the model
0
0
7
Now that Google has built some exceptional AI models, they only need to build application-layer products to use them like say, an RSS reader.
0
0
3
Are we heading towards a future where there will be more AI agents or AI investors?
1
0
2
“ultrathink. don’t give up. give it your all.” is good advice for humans too
1
0
14
Excited for Clay to counter the trend of AI companies opening Soho offices and instead open an office in the heart of San Francisco.
📍 We've just dropped our pin: in San Francisco! 🌉 Our 2nd office will be in the bay area and we're hiring! Want to hang with the team IRL and build great stuff while doing it? Apply for our open roles (in SF) here! https://t.co/4SPmbLeuLu
0
0
12
Something I'd love if someone built: I want a network-hosted UI/UX on Git Worktrees that I can SSH into, and have an Ngrok-like tunnel for local dev. Claude Code, Codex, and Cursor-Agent should be pre-installed.
1
0
7
I like GPT-5, and I think it makes a lot of sense to pare down the offerings in ChatGPT. That being said, there was something quite nice about the pristine professionalism of o3 that I'll miss.
1
0
9