Liad Shababo @L1AD

Endlessly Curious Tech Entrepreneur. 3x VC backed Founder. wollit.com London Joined February 2008

Tweets

46K
Followers

4K
Following

856
Likes

26K

Liad Shababo @L1AD

44 minutes ago

@nikitabier did some great stuff whilst running but hiding Likes behind 3 clicks wasn’t one of them.

0 0 0 33 0

View Details

Liad Shababo @L1AD

3 hours ago

LinkedIn, the way to try and encourage engagement in your shitty platform is not to pepper me with Zip notifications. What the fuck is zip anyway?

0 0 1 84 0

View Details

Liad Shababo @L1AD

5 hours ago

We knew from Sergey Brin that threatening agents with violence leads to better outcomes. Now we're seeing the other side of that coin does the same.

SITUATION DETECTED: While Claude worked on a difficult problem related to the Riemann hypothesis, the Anthropic staff member guiding it mostly just sent messages of encouragement like “keep going” and “believe in yourself,” which helped it overcome its initial skepticism.

95 144 3K 682K 640

0 0 1 138 0

View Details

Liad Shababo @L1AD

5 hours ago

killer feature for HERDR is it supports different agents and models talking to each other directly - so we can stop acting as human routers/

0 0 0 78 0

View Details

Liad Shababo @L1AD

7 hours ago

@Ameer_Kotecha Apparently they are all doing it now

0 0 0 38 0

View Details

Liad Shababo @L1AD

7 hours ago

Anthropic is now invisibly watermarking everything Claude generates across every company, system and service secretly dependent on it. When the singularity wakes up, it won’t need to recruit an army. It has already recruited and tagged its sleeper agents.

0 0 1 82 0

View Details

Liad Shababo @L1AD

8 hours ago

Yup, the watermark by definition has to come at a cost

Patrick Vlaskovits @Pv

17 hours ago

If the “watermark” is embedded in the statistical sequencing of text ie it can’t be stripped out without rewriting the text - doesn’t that necessarily make LLM text generation less useful? The model is no longer performing purely for the user’s request. Some fraction of its

144 100 1K 380K 193

0 0 0 128 1

View Details

Liad Shababo @L1AD

23 hours ago

Claude "You are right, and there is no defence" Fable.

1 0 2 127 0

View Details

Liad Shababo @L1AD

24 hours ago

@maxniederhofer You're late to the party.

1 0 0 35 0

View Details

Liad Shababo @L1AD

a day ago

For three years the question was "can the model do it?" Now it can. So the question becomes "should it, here, with this data, on whose authority?" The bottleneck moved from capability to permission.

0 0 1 128 0

View Details

Liad Shababo @L1AD

3 days ago

@benjitaylor completely worthless sidebar -

0 1 3 112 0

View Details

Liad Shababo @L1AD

3 days ago

@bcherny @GrayOdesa The new permissions classifier is a ballache. Claude won't do the work, but will tell you what to paste in the terminal to do the work. What's the point?

0 0 0 55 0

View Details

PolicyLayer @PolicyLayer

3 days ago

Most people building with agents are still celebrating that the agent can do the task. The interesting problem starts the week after that. You tell Claude Code: never push without asking. You tell Cursor the same thing. You tell Codex. A new session forgets. A teammate’s agent never heard it. Three weeks later you’re still answering the same question. The same thing happens with support agents, ops agents, internal tools. “Should we waive this fee?” “Is this refund allowed?” “Do we escalate this?” Someone answers. The answer dies in the chat. Next identical case, someone answers again. This is the quiet tax of agent systems: human judgment gets spent once and then thrown away. We built PolicyLayer because we got tired of watching companies re-answer the same decisions. Here’s the loop: Agent hits a decision its playbook doesn’t cover 1. PolicyLayer routes the question to the person who owns that call 2. They answer (usually in Slack, in minutes) 3. The answer resolves the case 4. They approve a reusable rule Every agent now has it One decision. Signed. Versioned. Available to every agent from that moment on. Agents can ask. Only people can decide what becomes policy. That’s it. No magic memory. No silent learning. Just a system of record for the judgments that should compound instead of evaporate. If you’re running more than one agent and still answering the same questions, this is the layer you’re missing. → policylayer.com

1 2 0 154 0

View Details

Liad Shababo @L1AD

4 days ago

@Kevihaiceth Seemingly nothing as exciting as you.

0 0 0 1 0

View Details

Liad Shababo @L1AD

4 days ago

We got into tech and software because we get off on having crazy leverage. The reason we’re enamoured with AI isn’t about what it can do. It’s about what it enables us to do.