VP Sales & GTM leader in B2B tech. Notes from the field on how AI is reshaping go-to-market: GTM strategy, GTM engineering, AI in sales. Sharp takes, no hype.
Hi, I'm Denis ๐บ๐ฆ AI enthusiast and lifelong learner, with years of experience in B2B sales in IT. I'm fascinated by how AI is changing the way we work and live. Here I share what I learn and the sharpest takes I find. If you're building with AI, let's connect and learn from each other. Follow along, DMs open ๐ค
Sunday in Seoul: Korea's finance minister, central bank governor and regulators held an emergency meeting before Monday's open. The market stabilization fund is being prepared, and buyback limits are eased for 3 months from Nov 24.
Doubts about AI spending are now a national policy problem, not just a Nasdaq one.
What would convince you the AI build-out is paying off: real revenue from AI products, or simply more capex?
@OptionKing666 MCP into Photoshop changes who the user is. Agencies that bill per asset will feel it first: the agent handles resizes and variants, while the designer keeps the brief and the final call.
@_ScottFriedman Matches the buyer side. Enterprises rarely pay for a model, they pay for something that already passed security review and a few renewals. ~$100M of revenue inside real workflows is proof no benchmark can give.
@stretchcloud Job posts often show the roadmap earlier than keynotes. "Proactive" is the word I'd watch: an agent that acts before you ask needs a lot of trust. Which permission do you think Apple hands it first?
A 2025 paper from PyMC Labs and Colgate is going viral again as "AI just killed market research."
What it actually shows: on 57 real product surveys (9,300 people), LLM "consumers" ranked concepts about 90% as consistently as real people agree with themselves on a retest. That's not "90% accurate."
I'd use it to drop weak ideas early, not to pick the winner. Would you trust it for that?
@misat0x Spanish-speaking audiences are huge and English creators mostly skip them, so this tracks. How do you check dubbing quality without speaking Spanish: a native viewer, or just retention numbers?
@xiaoba_ai Sales had this rule long before AI: a rep can write any email they like, but price and terms come from a system nobody improvises in. Agents built around the same split are much easier for a buyer to approve.
@bnchen Going from ~$800B to ~$2T a year means who lends starts to matter as much as who builds. Do you expect most of that gap to stay on the hyperscalers' balance sheets, or move to private credit and leasing deals?
@morganjingram Seen this pattern many times in B2B: months of emails go nowhere, then one hour off-script moves the deal. People buy from people they've actually met. A pickleball court beats step 9 of any sequence.
@RuslanVolkov25 "The AI writes the code. I decide what's true" is basically the job description now. With live deals in the mix too, which part still eats most of your week: building or selling?
49% to 74% with the same model. That's the number I'd show anyone picking AI tools by benchmark alone. When you compare agent products, do you ever test the plumbing (which tools get loaded, how much comes back) or just the model name? https://t.co/hlZYeBPHql
This Anthropic tool use paper is f*cking gold
Anthropic's own team took Opus from 49% to 74% accuracy and cut token use by 85%. And the model wasn't the upgrade.
The upgrade: they stopped feeding Claude every tool and every raw result.
Tool search โ defer loading โ code calls โ parallel โ examples
The 6 parts of advanced tool use:
Part 1 โ Tool search
58 tools ate ~55,000 tokens before the agent read a word. Search loads ~500 tokens and finds the 3-5 it needs
77,000 โ 8,700 tokens, and on their MCP evals Opus 4 went from 49% to 74%
Part 2 โ Defer loading
Every tool marked defer_loading, only your 3-5 most used stay loaded
Search "github" and only the GitHub tools show up
Part 3 โ Call tools from code
One script reads 2,000+ expense lines in a sandbox
The model only sees the 3 people over budget. 200 KB โ 1 KB
Part 4 โ Run in parallel
20 tool calls in one code block, 19+ inference passes skipped
43,588 โ 27,297 tokens
Part 5 โ Show examples
The schema says "string". The example says "2024-11-06"
Accuracy on complex parameters: 72% โ 90%
Part 6 โ Fix the bottleneck
Bloat โ search. Big results โ code. Bad params โ examples
Opus 4.5 went from 79.5% to 88.1%
More tools in context made the agent dumber. Fewer tokens made it smarter.
I packed all 6 parts into one prompt. Drop it into Claude Code and it sets the whole tool stack up for you.
6 parts. 1 prompt. 49% โ 74%.
Then read the article below โ
@Anujatk14 Agree, shared prompts don't transfer well in creative work: same prompt, different shot every time. A 3D shot composer sounds like the right direction. Do people use it to plan shots before generating, or to fix them after?
@ganesh2412@X That split makes sense: QA proves the evidence holds up, release and security owners sign off on adoption. Good luck with the AgentProof pilots.
@kokkisajee Fair point. If MCP and agent hooks give extensions the same reach, the fork is mostly a head start. The part that might last is usage data: seeing which edits people accept or throw away. Can an extension ever see that as well as the editor owner?