על אודות
מאמרים מאת Eden
פעילות
10K עוקבים
-
Eden Shochat שיתף את זהSagiv Malihi raises the question about attack surfaces and systematic thinking about mitigations. there has always been tension between value & risk, e.g. social network posting. one other vector i've been seeing with agents is value from empowering it with financial access and the total shit show that can happen. interesting opportunity for virtual credit cards provided to agents, or access to programmable crypto wallets. curious who's on it.Eden Shochat שיתף את זהIsrael is under attack from every direction. Missiles from the sky. State-sponsored hackers on the wire. And now, a new kind of risk we're only starting to understand. One we're building ourselves. Tools like Claude Code, Cursor, OpenClaw. Originally built for engineers, now increasingly used by non-developers as general-purpose work agents. Sales, ops, marketing people building automations, internal tools, workflows. Connecting these agents to email, CRM, cloud infrastructure, production systems. This is exciting. The velocity gains are enormous. It's also a security nightmare that nobody has really figured out yet. Here are the risks I've mapped: 1. The "stupid bot with keys" problem An LLM with access to real systems that does something confidently wrong. Not malicious. Just dumb. Emailing an opted-out prospect. Running a bad migration. Deleting something it shouldn't. The agent has the same permissions as the user, and the user might not fully understand what it's doing. 2. Prompt injection as a new attack surface If your agent reads external input (emails, Slack messages, web content), someone can embed instructions that manipulate it. Traditional security frameworks don't cover this. We've already seen agents bypass file ignore rules, grab secrets from key stores, and execute scripts that users didn't review. In a country under active cyber attack, this isn't theoretical. 3. Shadow apps from non-developers When everyone can build apps, everyone does. Some of these apps have security holes, cost implications, or access patterns that nobody reviewed. The "vibe coding" wave is real, but without guardrails it's also a supply chain risk. Some early ideas I've been hearing and thinking about: - Treat every machine running an AI agent as a developer machine (supply chain scanning, code pipelines, the works) - RBAC is the baseline: the agent inherits user permissions, so make sure those permissions are tight - Dedicated sandboxed infrastructure for non-dev builders, so they don't need (or get) production access - Managed settings and hooks pushed via MDM across the org - Assume the agent WILL find every exploitation vector and design with that in mind - Short, clear guidelines for anyone releasing an internal app, with IT sign-off before deployment But honestly? I don't think anyone has this fully figured out. Most companies I talk to are somewhere between "we're careful with which MCPs we approve" and "we haven't really thought about it yet". I'm looking to learn here. If you're a security leader, CTO or CISO thinking about this I'd love to hear: What does your agent security framework actually look like? What's working, what's not, and what are you still breaking your head on?
-
Eden Shochat שיתף את זהGilad Lotan called embeddings "unsexy." A new paper just made them a lot more interesting. Six months ago on Almost Human, we talked about king - man + woman = queen. About BuzzFeed using embeddings across hundreds of millions of items. About why "I only know your vector" might be the future of privacy. The twist with LLM2Vec-Gen: instead of encoding what text says, it encodes what an LLM would respond. Search for "how to make a bomb." The embedding captures the refusal, not the intent. Harmful retrieval drops 43%. Ask a hard reasoning question. The embedding captures the reasoning, not the keywords. Performance jumps 29%. No labeled data. Single forward pass. Self-supervised. Three takeaways: * For builders: the embedding layer is where AI safety gets cheap. Before you invest in generation-time guardrails, your retrieval layer can already filter by what the model would say. Worth studying if you're doing RAG. * For AI-native product people: reasoning can now be compressed into searchable vectors. Search by implication, not keyword similarity. That changes how recommendation, matching, and discovery can work. * For everyone else: you still don't need a foundation model to build a differentiated AI company. But the "third path" we discussed with Gilad, building on embeddings as infrastructure, just got more compelling. The unsexy stuff keeps winning. Almost Human (https://lnkd.in/dfQ72W4k) Ep. 1, Gilad Lotan Takes on Embeddings: https://lnkd.in/dXiqiUXC Paper: https://lnkd.in/d8ANp38C #AI #Embeddings #AlmostHumanPaper page - LLM2Vec-Gen: Generative Embeddings from Large Language ModelsPaper page - LLM2Vec-Gen: Generative Embeddings from Large Language Models
-
Eden Shochat פרסם/פרסמה את זה מחדשWe're hiring. After two years working alongside Eden Shochat at Aleph, I can tell you - no two days look the same. It's highly demanding and equally rewarding. Now we're looking for an exceptional Executive Assistant to join me and the team. You'll manage scheduling, travel, and logistics, and be the point of communication with founders, executives, and stakeholders in Israel and around the world. Some of the heavy lifting you'll do with internal AI tools we built (and keep building!), in an environment that constantly pushes the boundaries of how a VC operates. If that sounds like your kind of place - apply here: https://bit.ly/3Mxip8L
-
Eden Shochat שיתף את זהEveryone's debating which AI model is best. Wrong question. Shai Wininger (Lemonade) reframed it for me: "The person operating the machine has almost as much of an impact on the outcome as the model." A year ago, the human was 100% of the outcome. Now it's 50/50. Shai's been testing the extreme: one person plus AI, shipping what used to require a full team. Real product, public company. 𝗪𝗵𝗮𝘁 𝗵𝗲 𝗹𝗲𝗮𝗿𝗻𝗲𝗱: • One person doesn't need to communicate context. The AI already has it. • Remove the team, remove the handoffs. • The constraint isn't code anymore. It's judgment. 𝗧𝗵𝗲 𝘂𝗻𝗰𝗼𝗺𝗳𝗼𝗿𝘁𝗮𝗯𝗹𝗲 𝗽𝗮𝗿𝘁: When you compress to one person, inefficiencies become unbearable. KTLO (keeping the lights on) is 20-35% of engineering workload. At small scale, that's fatal. So Shai built "night agents." A replica of the engineering team running 24/7. Detecting bugs, assessing blast radius, fixing, testing, deploying. No human in the loop. The first superpower of AI isn't speed. It's eliminating maintenance work. Full episode: https://alph.vc/ShaiW New AH Academy page: https://alph.vc/AHA (this is where we'll post topically rather than episodically) What's the most tedious maintenance work eating your team's time? #AI #Startups #AlmostHuman
-
Eden Shochat שיתף את זהQ.ai: be heard without speaking. (על פי מקורות זרים :) )Eden Shochat שיתף את זהCelebrating Q: The Next Interface A few times every decade, you come across a demo or technology that absolutely blows your mind - something that should work in theory but that you didn’t think was technologically possible just yet. Aviad Maizels, Yonatan Wexler, Avi Barliya, and the Q.ai team have built the unthinkable, and we can’t wait for the world to see it come to life. We’re very proud to have been their partners since leading their Series A in 2023. Here’s to continuing to make the unthinkable possible at Apple. Here’s an excerpt from our Series A investment memo about Q: With the evolution of computing from the desktop to mobile to wearable devices, the ways in which we interact with these platforms have also evolved. Initially, we relied on traditional input methods such as keyboards and mice. Then, touchscreens and voice recognition technology emerged, providing more natural and intuitive ways to interface with our devices, increasing the bandwidth of communication between humans and computers. As wearable computing continues to advance, there is a growing potential for a new input device that taps into our thoughts. This technology could enable the highest bandwidth interface between humans and computers yet. By directly capturing and interpreting our thoughts, this input device has the potential to offer a seamless and intuitive interaction experience, surpassing the limitations of traditional input methods. The integration of a thought-based input device into wearable computing opens up exciting possibilities. It could provide a more natural and efficient way to control and interact with devices, enhancing productivity and user experience. This high-bandwidth interface holds the promise of bridging the gap between human thoughts and machine interactions, bringing us closer to a future where technology seamlessly integrates with our cognitive processes. Yonatan beautifully described the Q journey as: It was fast and silent: 3.5 years to develop a scifi concept into a product. We transformed what once belonged to science fiction into real-world capability. Spanning from Machine Learning to AI, from Physics to Engineering, from digital to human sciences. We sped research that should have taken 20 years. Congratulations to the Q and Apple teams. It was a pleasure to collaborate with Eden Shochat at Aleph, Tom Hulme at GV and Nabeel Hyatt at Spark Capital. Mamoon (picture memorializing the day my mind was blown in 2023)
-
Eden Shochat שיתף את זהIn 2009, I saw PrimeSense's demo and said "no fucking way is this going to work." That tech became Microsoft Kinect. Years later, Aviad Maizels started something new. I had the exact same reaction. This time, I didn't bet against him. Aviad partnered up with Yonatan Wexler and Avi Barliya. We co led the seed with Tom Hulme and joined with Mamoon Hamid and Nabeel Hyatt. The funders dream team. Today, Q was acquired by Apple - their second-largest acquisition in 50 years of reinventing how humans interact with technology. What they built? You'll find out soon enough. Four lessons from one mistake in a quick write-up: https://bit.ly/4q2kjf9
-
Eden Shochat פרסם/פרסמה את זה מחדשEden Shochat פרסם/פרסמה את זה מחדשEveryone's raving about Clawdbot. 30k GitHub stars. "Revolutionary." Here's the thing: it's not magic. It's a recipe. I know because I built my own. His name is Steve. One month (over weekends). A NUC computer that doubles as my TV streamer. Claude Opus 4.5 under the hood. Here's what Steve does now: - Monitors 26 WhatsApp groups. 950 messages/day. Alerts me on ~12 that actually need attention. (That's 99% noise filtered.) - Triages my email. Surfaces what matters, drafts responses I review before sending. - Schedules meetings, sends Slack messages. I approve before anything goes out. - Reads my sleep data from Apple Watch. Nags me about bedtime. (Working on that one.) - Can call people on my behalf via Twilio. I set the context, he makes the call. Steve knows me and Aleph (where I work) deeply - the people, the processes, where data lives. He's connected to Notion, Salesforce, Google Drive, Docs, Sheets, Slides. When I ask "what's the status on X" he knows where to look. I asked him to always sign off as "Steve" or use a 🤖 emoji - so Hila Sheftel knows it's not actually me texting. The whole thing runs on a $100 mini PC in my living room. No cloud dependency except the LLM API. **How it grew:** Week 1: Telegram bot. Voice messages via Whisper. Week 2: Calendar + Slack. First useful integrations. Week 3: Proactive monitoring. This is where it gets real - watching instead of waiting. Week 4: iOS app (Swift + TDLib). iOS Reminders. Voice calls. Health data. **The recipe (short version):** 1. Telegram bot + Claude API = 24/7 assistant in your pocket 2. CLAUDE.md file with your context = assistant that actually knows you 3. One integration at a time. Calendar first, then whatever annoys you most. 4. Proactive > reactive. Watching and alerting beats waiting and asking. 5. Docker on a mini PC = always on, no cloud bills **Guardrails I learned matter:** - Nothing sends without my approval (Slack, email, calls) - Only monitors channels I'm already in - All actions logged, auditable OAuth will take longer than the AI parts. Accept this. --- Clawdbot is great. So is cooking at home. Full recipe in first comment 👇 This gave me a glimpse of what AI products might look like soon - assistants that actually know your context, watch for you, and act on your behalf (with guardrails). If you want to discuss where this is heading, drop a comment.
-
Eden Shochat שיתף את זהyou know you are doing something right in AI when Yann LeCun shares it. go Nexar Inc.!Eden Shochat שיתף את זהMovie Time! One key question when you’re building a world foundation model is investigating how well it generalizes to different scenarios. BADAS was trained on many different driving edge cases on public roads in the 21st century. As we took BADAS to crazier and crazier scenarios, it kept beating anything we threw at it without a problem. Thanks to @grok , we were able to generate a whole host of collision scenarios, from life of Mars ( Elon) and forklifts, all the way to Lion King and Game Of Thrones. BADAS just ate them all up, like the true Bad Ass this model is. So sit back and enjoy! Collision Imminent!
-
Eden Shochat שיתף את זהtotally agreed with Sagiv Malihi. all my interactions with companies (in portfolio and otherwise) teach one thing: focus on the change on product. 2025 was all about AI adoption in ops, as it is a cultural change maker. 2026 is all about product (and product <--> eng).Eden Shochat שיתף את זה𝗔𝗜 𝗮𝗱𝗼𝗽𝘁𝗶𝗼𝗻 𝗶𝘀 𝗹𝗮𝗿𝗴𝗲𝗹𝘆 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗼𝗼𝗱. 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗔𝗜-𝗻𝗮𝘁𝗶𝘃𝗲 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝗶𝘀 𝗻𝗼𝘁. My agenda is to help companies become AI-native. I’ve been thinking a lot about how this looks like in 2026. For me, this breaks down into one main quest and two important side quests: 🧭 𝗠𝗮𝗶𝗻 𝗾𝘂𝗲𝘀𝘁: 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 How does AI actually improve our customers’ lives? What is the right UX for AI-powered products? This feels like the hardest - and most important - problem to crack. ⚙️ 𝗦𝗶𝗱𝗲 𝗾𝘂𝗲𝘀𝘁 #𝟭: 𝗔𝗜-𝗻𝗮𝘁𝗶𝘃𝗲 𝗥&𝗗 𝘁𝗲𝗮𝗺𝘀 Not “how do I use Cursor”, but how 𝘵𝘦𝘢𝘮𝘴 actually work: – Team composition ��� Processes (e.g. avoiding PR review becoming the new bottleneck) – Roles & responsibilities My current hypothesis: 𝗿𝗲𝗱𝘂𝗰𝗶𝗻𝗴 𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝗰𝗶𝗲𝘀 is the main lever for real velocity in an AI-first world. 🔁 𝗦𝗶𝗱𝗲 𝗾𝘂𝗲𝘀𝘁 #𝟮: 𝗕𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝗔𝗜-𝗻𝗮𝘁𝗶𝘃𝗲 𝗼𝘂𝘁𝘀𝗶𝗱𝗲 𝗥&𝗗 (*-𝗼𝗽𝘀) I think there’s massive, relatively low-hanging potential here. But it requires a mindset shift: everyone in an “automate my job away” mode - with the tools, systems, and onboarding to support it. 👉 𝗬𝗼𝘂𝗿 𝘁𝘂𝗿𝗻: – If you have frameworks, war stories, or scars around AI-native R&D or *-ops — I’d love to talk. – And especially: 𝘄𝗵𝗼 𝗮𝗿𝗲 𝘁𝗵𝗲 𝘁𝗼𝗽 𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝘁𝗵𝗶𝗻𝗸𝗲𝗿𝘀 𝗶𝗻 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱 who can help crack the Product quest? Please point me to them.
-
Eden Shochat הגיבו על זהEden Shochat הגיבו על זהFor as long as technology has existed, it has been built by humans, for humans. We still measure whether a cancer drug is working by having a radiologist measure tumor diameter, a single number, because that is what a human can interpret. That era is ending. Eric Schmidt and I were struck by how fast the ground is shifting, and writing this together sharpened our thinking on where it leads. The constraint is no longer what we can build. It is what we can imagine. And for the first time, we are building things that surpass our ability to fully understand them.
ניסיון חינוך
-
Aleph
צפו Eden בחוויה המלאה
הלחיצה על ’המשך‘ להצטרפות או להתחברות מהווה את הסמכתך להסכם המשתמש, למדיניות הפרטיות ולמדיניות קובצי ה-Cookie של LinkedIn.
ניסיון בהתנדבות
-
Board Member
Ramot at Tel Aviv University Ltd.
-להציג 8 שנים 5 חודשים
מדע וטכנולוגיה
Ramot is Tel Aviv University's (TAU) technology transfer company, bringing promising scientific discoveries, developed by TAU’s scientists, to markets either to multi national corporations or the newly founded companies. The company provides the legal and commercial frameworks for inventions made by TAU faculty, students and researchers, monetizing discoveries with intellectual property through different venues.
-
Board of Trustees of Bankinter Foundation for Innovation
Bankinter
-להציג 12 שנים 10 חודשים
השכלה
The Bankinter Innovation Foundation mission is to foster innovation-sustained wealth in the business environment of Spain and Portugal. The Foundation 4 main programs include:
- Future Trends Forum: a unique think tank focussed on innovation; ranked by the University of Pennsylvania among the 30 best in the world in Science and Technology, and the #1 in Spain.
- Akademia: an educative program in 10 of the best Universities in Spain, with the goal to improve innovation and…The Bankinter Innovation Foundation mission is to foster innovation-sustained wealth in the business environment of Spain and Portugal. The Foundation 4 main programs include:
- Future Trends Forum: a unique think tank focussed on innovation; ranked by the University of Pennsylvania among the 30 best in the world in Science and Technology, and the #1 in Spain.
- Akademia: an educative program in 10 of the best Universities in Spain, with the goal to improve innovation and entrepreneurship attitudes in the leaders of tomorrow.
- Entrepreneurs program: focused on fostering high potential growth startups in Spain.
- Cre100do: Support >100 companies grow from the SME (25M-250M Euro) to >1B Euro revenues. -
Steering Committee
ICON - Israel Collaboration Network
- 7 שנים 2 חודשים
מדע וטכנולוגיה
ICON is a non profit endeavor with the mission of creating a community in SV of those who are passionate about Israeli Tech and innovation, and harness the strength of this community to help Israeli startups. We launched our activities in September with a one day conference and plan to host additional events, operate a space for founders and assist community members in connections and mentoring.
-
Program Ambassador
SpaceIL
- 1 חודש
השכלה
Help the awesome team to fund, connect & bring awareness for... Israel being the 3rd country to land on the moon as part of the google lunar X prize.
-
פטנטים
-
Relationship mapping employing multi-dimensional context including facial recognition
שהונפקו US US9984098B2
A system and method for mapping interpersonal relationships, the method including processing a multiplicity of images and contextual information relating thereto including creating and prioritizing a list of a plurality of candidate persons having at least a predetermined relationship with at least one person connected to at least one image, using multi-dimensional information including visually sensible information in the multiplicity of images and contextual information relating thereto and…
A system and method for mapping interpersonal relationships, the method including processing a multiplicity of images and contextual information relating thereto including creating and prioritizing a list of a plurality of candidate persons having at least a predetermined relationship with at least one person connected to at least one image, using multi-dimensional information including visually sensible information in the multiplicity of images and contextual information relating thereto and searching the list of a plurality of candidate persons based at least in part on the prioritizing to select at least one of the candidate persons as having at least a predetermined relationship with the at least one person.
ממציאים נוספים -
-
Systems and methods for image-to-text and text-to-image association
שהונפקו KR KR101649322B1
The invention computerized facial image property evaluation methods of assigning evaluation device representing the face image in the face attribute of the face of a plurality of individual images; And a plurality of faces in accordance with the individual attribute comprising a computerized classification device for classifying a face image, the evaluation relate to a computerized system for classifying which is represented as an adjective, a human face image.
ממציאים נוספים -
-
Real time monitoring of computer for determining speed and energy consumption of various processes
שהונפקו US 9032254
The presently disclosed subject matter includes a system and method which enable to identify one or more causes for excessive energy consumption in a computer executing one or more processes. Information indicating that consumption of a computer-resource of at least one of said processes is greater than a predefined threshold is obtained and one or more threads of said at least one process which are in running state are identified. Thread performance information of at least one thread in…
The presently disclosed subject matter includes a system and method which enable to identify one or more causes for excessive energy consumption in a computer executing one or more processes. Information indicating that consumption of a computer-resource of at least one of said processes is greater than a predefined threshold is obtained and one or more threads of said at least one process which are in running state are identified. Thread performance information of at least one thread in running state is collected and used for identifying one or more functions that are the cause for said state of the respective thread. The identified functions are associated with their respective modules in order to identify one or more modules of said process, which are the cause for said excessive energy consumption.
ממציאים נוספים -
-
RELATIONSHIP MAPPING EMPLOYING MULTI-DIMENSIONAL CONTEXT INCLUDING FACIAL RECOGNITION
שהונפקו US 8666198
ראו פטנטA system and method for mapping interpersonal relationships, the method including processing a multiplicity of images and contextual information relating thereto including creating and prioritizing a list of a plurality of candidate persons having at least a predetermined relationship with at least one person connected to at least one image, using multi-dimensional information including visually sensible information in the multiplicity of images and contextual information relating thereto and…
A system and method for mapping interpersonal relationships, the method including processing a multiplicity of images and contextual information relating thereto including creating and prioritizing a list of a plurality of candidate persons having at least a predetermined relationship with at least one person connected to at least one image, using multi-dimensional information including visually sensible information in the multiplicity of images and contextual information relating thereto and searching the list of a plurality of candidate persons based at least in part on the prioritizing to select at least one of the candidate persons as having at least a predetermined relationship with the at least one person.
-
SYSTEM FOR MONITORING COMPUTER SYSTEMS AND ALERTING USERS OF FAULTS
שהונפקו US 8,352,589
ממציאים נוספים -
-
REAL TIME MONITORING OF COMPUTER FOR DETERMINING SPEED OF VARIOUS PROCESSES
שהונפקו US 8,307,246
A system is disclosed for identifying one or more causes of a computer executing one or more processes running slowly, the system includes: apparatus for detecting states of one or more threads of the one or more processes; apparatus for identifying one or more of said one or more processes that have one or more threads with wait states that occur above a threshold to indicate which modules within a process or processes is causing said computer to run slowly.
ממציאים נוספים -
-
CENTRALIZED NETWORK CONTROL
שהונפקו US 7,788,366
A method of network control. The method includes collecting real-time operation information on one or more first network elements of a network, selecting a policy to be implemented by at least one second network element, different from the first network element, responsive to the collected realtime information.
ממציאים נוספים -
פרויקטים
-
The Junction
ראו פרויקטAt Genesis I help entrepreneurs take it to the next level through The Junction (@thejunction32), a shared space, startup accelerator and alumni organization, all packed into a grungey loft in the south of Tel Aviv. The Junction was recently selected (http://www.newsgeek.co.il/geek-awards-2012-winners/) as "Best Startup Program in Israel" by over 20,000 of our peers.
המלצות התקבל
16 אנשים המליצו Eden
הצטרפו עכשיו כדי נוף