Passionate about tech and how it can improve lives. Founder @PubNative, ex Co-CEO @VerveGroupHQ Building next-gen marketing agents at https://xitter.cf/t.co/5JFeKgfVn6
Berlin
Joined February 2009
- Tweets9.1K
- Following4.9K
- Followers1.5K
- Likes75.1K
Pinned Tweet
I've spent now - what? - 15 years around ads? a long time.. Longer than I expected.. esp. since I didn't target it specifically and I didn't realize it was an ecosystem at start.
My first adserver was open source and build in PHP in the early 2000s.
Those were the days...
Such a great event! and a very cool location also!
A couple more pics from the workshops and the demo; in no particular order.
Didn't finish my project... but the weekend is coming :D
Roughly 24 hours ago we concluded the Berlin @OpenAIDevs DevDay Hackathon!
Over 245 people joined us on two floors in the @circuit_ber startup house.
Within 8 hours participants had the chance to build amazing projects, no matter if technical or non technical.
In the end, 73 projects were submitted across the Developer and Founder/GTM tracks.
3 teams per track won a lot of API & Codex credits. As well as 10 tickets to the DevDay Berlin tomorrow!
For me it honestly was a huge amount of stress and hassle!
I straight up ran around the location from 8 am to 7pm and tried to put out as many fires as possible.
We had a lot of typical hackathon things go wrong as usual haha
Wifi was spotty and going down, credit distribution was flaky but in the end we managed to persevere.
I was lucky to have an amazing team of supporters on location that was pitching in and honestly handling so much work that I could've never done without them.
This day showed how important real life interaction still is and how much I personally still can learn about making this experience enjoyable for everyone.
Thank you to the brilliant ambassadors who were there:
@DFinsterwalder (from Göppingen)
@mrlesk (from Vienna)
@laraavci_ (from Munich)
@janpfrenger (from Frankfurt)
@artur_wala (from Warsaw)
as well as the wonderful extra supporters:
@_Ms_Stacy, Eleni Koukidou, Jérémy Grandillon
and tons of other people who helped out on the event
Special thank you also to the absolute legendary OpenAI folks who showed up:
@paw_lean: for making this event possible in the first place and being a beam of inspiration for everyone on the event
@reach_vb: for starting this community program and also showing up at this event!
and @TilmanResch for dropping by for a few minutes and being the OpenAI GTM legend he is
And thank you to @rhiannon_io who unfortunately couldn't make it to the event but helped immensely to actually let this come together and deliver goodies for so many people from the US to Europe.
Thank you to @Qu1nn_web3, Kushal Singh and the rest of the location staff for managing the different event fires with me and making sure the whole event is well caffeinated.
And thank you to @Son_of_a_Pigeon, the best event photographer ever, for the most insane and awe-inspiring pictures from the day.
To even greater hackathons and more insane projects
Max
PS:
Excited for the OpenAI DevDay Berlin tomorrow!
ALT happy but exhausted event team
Great breakdown!
And yes every other question (eg auth) goes downstream from here;
If I may add one thing - the two specs seem rather complementary to me (more from the use-case perspective):
- for example if you have an MCP App dashboard - you could control it (filters, date pickers) via WebMCP - but if you want to change the data behind it - that could be an MCP tool call
- one step further: for multiplayer this is even more interesting; different agents can work on a canvas on behalf of different users via webMCP but save state or retrieve context via MCP (without disturbing the canvas)
this graphic is way more confusing than it needs to be. tldr:
- WebMCP: tools defined in the JS of a page, therefore can only be called when that page is open in a browser.
- MCP: tools defined on a server that can be accessed via HTTP, therefore can be called from anywhere.
if you understand this, you understand all of the follow-up implications around auth, UI, etc.
Probably the keynote I was most excited about? Couple of points:
1. Conversations using plugins grew ~10x this year - is this the new "appstore interaction model"?
2. Don't use tool search or hide tools behind list_command + run_command. expose every tool.
ChatGPT handles discovery, and admins get per-tool permissions
3. Keep schemas lean. ~3 in 4 tools are under 1kB. the biggest hit ~260kB, and when a schema is too big the model only sees "args: object" and loses the param guidance
4. Annotations are required for plugins. verified read-only tools get screened less and run faster - we've definitely been battling with this one :)
5. Only ~56% of servers use server instructions. keep them to ~500-1,000 chars, most important lines first
Last but not least (in the post prezi interview) -> @sarfata described some usecases for MCP events -> one of which being collaborating with agents on the same "canvas".
Such an unlock!
Looking forward to the demos later!
For those who missed it I took some pictures (cropped them with Codex); sharing them below 👇
@sarfata from @OpenAI shared a key takeaway during his “MCP in the Wild” keynote at AGNTCon + MCPCon NA: Make every tool discoverable and ensure annotations accurately describe its capabilities.
Because an agent can't reliably select the right tool if it doesn't know what that tool can do. Livestream: bit.ly/3U6tMrS
In the interim - a prerequisite of of getting there - is that Anthropic and OpenAI (and similar others) become “platforms” - meaning the value they create across their ecosystems and the broader economy (consumer and enterprise) is higher than the value they capture = own revenues (and ideally valuations also).
Given how things are accelerating it seems only a matter of time.
this was GREAT!
I quite like the platform product strategy part. I think Sam is somewhat right in regards to missing the iPhone moment with how people interact with AI. And they did somewhat place a bet on OpenClaw.
At the same time, I spend most of my time in agents now rather than anywhere in the browser (using ChatGPT more and more, and the Chrome plugin).
I do think there are two angles here: B2B and B2C. They have different trajectories and value cycles, and this technology impacts both but in different ways (some consumer usecases move, some new ones are created.. ). The separation between product (ChatGPT / Codex) and API, even as those come closer together, I feel more of a B2B lean on the API and maybe a 50/50 B2B and B2C split for Codex? (though again, I use it much more for work).
That said, I like the point about not competing with customers. I hope that holds long term (we have seen a couple of diverging moves form other large corps).
My main point is that LLMs and models need to generate meaningful, monetizable value: not just cost decrease, doing things faster, or fewer hours, but actual concrete value producing (we see signs of it already both at the higher end and lower end).
If new businesses are being built, they should push GDP (even if slower due to inertia).
A couple of platforms have managed to achieve this.
Bill Gates had a definition for this at Microsoft: basically, the value of the businesses built on the platform is larger than the value of the platform itself.
And there are some examples: Google with web publishers, Shopify with ecomm and Apple with the App Store. I am not sure if the total value is larger for the latter but there was definitely a virtuous cycle with developers building on top of the Apple ecosystem, getting distribution, and most importantly, monetization.
So maybe monetization is the piece that is missing?
My conversation with Sam Altman (@sama), co-founder & CEO of @OpenAI.
0:00 Tobi Lütke, AI-Native Companies & Why Adoption Moves Slowly
5:45 Sam's Own Resistance to AI & the Missing iPhone Moment
10:00 Models, Compute, Power Laws & Non-Consensus Talent
18:37 From AI-Obsessed Kid to Founder, Investor & Back Again
23:19 Impossible Problems, Scientific Discovery & Human Connection
30:16 AI's Two Biggest Risks: Loss of Control & Centralized Power
33:09 Iterative Deployment, AI Safety & Learning From Reality
40:27 Why People Fear AI & the Coming Small-Business Boom
46:13 Context, Memory & the Next Way We Will Work With AI
49:17 OpenAI's Platform Strategy & Killing Good Ideas
53:20 Peter Thiel, Paul Graham & the Value of Nonlinear Thinkers
1:00:42 How Y Combinator Changed Startups & Shaped OpenAI
1:04:03 Learning More From Success & the Power of Repetition
1:09:45 Building OpenAI Without Customers, a Product or a Playbook
1:15:57 Letters to His Son & Preserving the Story
Includes paid partnerships.
I realized, than when you type, you synthesize more, and obviously there's no babbling and you're more concise.
While when you dictate, there are filler words, and there's basically more content, so the signal-to-noise ratio is lower.
Obviously, this is a dictated post. And I am beginning to recognize dictated posts as much as I recognize AI-written ones.
The ouroboros starting to eat its own tail - barchart.com/story/news/3584…
Does this mean I can now ask ChatGPT to give tasks to Claude? youtu.be/_Gd9yzAc-WI?si=0R9s…
Should I build a mobile app for this?
The dimensions seem to fit..
(Touch) controls + lights should not be that hard.. even the fidget joystick is doable
the only challenge could be the mapping to actions but worth exploring
Meet kbd-1.0-codex-micro, built with @work_louder.
Map the buttons and joystick to your workflow, and keep your pinned chats in view.
Get yours before stock returns 410.
The new AppStore is winking at me.
With tales from its forefathers...
This was in Claude Web mind you..
#crosspromo
Claude said they won't do ads.. but if they do ads - it's gonna be for their own products
and from their products.. they picked the one that's not available..
so that we don't forget about it.. until it's available again?
is that.. the definition of a fable?
Claude can run ChatGPT ads.
Well, mostly read them, for now. But that’s where this is heading.
A couple of weeks ago ChatGPT took their ads platform out of beta and opened it to the public. With it, they opened up not just the ads manager but their advertiser API too.
We’re somewhat of a Claude shop at hypd.ai, so we couldn’t help but think: can we run them in Claude? The answer was: not directly.
So we used Claude Code to build an MCP Server and bridge that gap. The model context protocol is open source and supported by both Anthropic and OpenAI.
PS: I was having a bit of a deja-vu writing this — a couple of months ago I was joking about the Claude Ads Agent in Claude Code.. and now… here we are :)
Some more details on ChatGPT Ads, the process of building the MCP and what is next:
linkedin.com/pulse/how-we-br…
out of Beta 🎉
we also offer a demo account with dummy data, if anyone wants to "try before you buy"
don't hesitate to DM, happy to enable it
This is one framing; but the framing I prefer is around value creation;
One engineer with AI = 3x value creation than before (just an example) - for existing customers:
One sales ops = 5x value creation for client facing roles.
What is the limiting factor for revenue generation? If it’s humans, or the GDP - you can optimize everything underneath but you can’t change reality, at least not that fast (some CEOs can well distort it though). And you depend on other “cogs” of the economy.
Tl:dr efficiencies are measured against the harsh reality of the P&L which is basically the “laws of physics” for a business.
what about sleeping? :D
I can hardly sleep since Opus 4.6 in Claude Code; #shippinganxiety
