Instagram story viewer> @yishaiback> Posts

Yishai Back

@yishaiback

149

followers

261

following

Never let your team solve the same problem twice, with Omnishi
The durable memory layer for AI, what Microsoft CEO calls "the new IP of the firm"

POSTS STORIES REELS TAGGED

Download All

Israel is seeing both sides of the AI boom: six-person startups selling for $100 million, while major employers prepare to hire fewer people.

Geniez AI, a startup with just six employees, was reportedly acquired for up to $100 million. Meanwhile, Israel Aerospace Industries is slowing recruitment as it evaluates how AI could reduce future staffing needs.

What’s striking is the union’s response. Instead of trying to block the technology, workers’ union chairman Yair Katz has hired an AI adviser and is investing in training, courses, and skills adjustments for employees.
His message: resisting technological change may delay it, but preparation is the better long-term strategy.

That contrast will increasingly define the AI economy: small teams building extraordinarily valuable companies at record speed, while large employers reconsider how many people they need.

Capital is adapting quickly. The real question is whether organizations can help workers adapt just as fast. by @yishaiback

3

0

7 hours ago

AI was allegedly used to help build a covert recorder. Then AI was used to help trace it. 🤖🔍

OpenAI discovered a recording device hidden behind a television in an executive conference room at its San Francisco office. Federal prosecutors allege that OpenAI engineer Abdulrahman Alfozan installed it after an account linked to him asked a chatbot about continuous recording, Wi-Fi connectivity, audio quality and encrypted storage.

OpenAI says its own models helped investigators understand the device, trace its origin and identify the employee. Alfozan was subsequently fired and arrested by the FBI.

Security teams increasingly need to connect physical access, device forensics, internal tool activity and AI-assisted investigation. The same technology that can lower the barrier to planning an operation can also accelerate its detection and reconstruction.

OpenAI says it found no indication that model weights, source code or customer data were accessed or taken. by @yishaiback

2

0

8 hours ago

Israel has been trending on X. Not because of politics, but because visitors keep noticing two things: cranes and children. 🇮🇱

Posts comparing Israel’s streets and skyline with those of other Western countries have drawn hundreds of thousands of views. The same observations appear repeatedly: new towers rising across Tel Aviv, construction continuing despite the war, and families filling public spaces.

The data helps explain why the contrast feels so striking.

Israel remains the only high-income OECD country with fertility above the 2.1 replacement level, at roughly 2.9 children per woman. Its population has passed 10.3 million, driven primarily by births rather than immigration.

At the same time:
🏗️ Housing construction has remained elevated despite labor constraints
📈 GDP grew by approximately 3% in 2025
💻 High-tech exports reached record levels
🌍 Foreign investment remained strong

Most wealthy countries are confronting the same combination of declining birth rates, aging populations, and slower growth. Israel is moving in a different direction.

That does not make it immune to demographic or economic pressures. But it does make Israel an increasingly important case study: a developed economy where family formation and continued building coexist.

The wider lesson may be the most interesting one. Demographic decline is often discussed as though it were inevitable. Israel suggests that it isn’t. by @yishaiback

4

0

13 hours ago

🚀 Microsoft is making its biggest bet yet on custom AI chips.

CEO Satya Nadella says Microsoft is preparing to scale its own Azure Maia processors to multi-gigawatt deployments starting in 2027.

The goal? Make Microsoft's own chips the primary infrastructure powering AI across products like Teams and Copilot.

⚡ The efficiency gains are significant:
• Maia 200: Up to 1.7x more useful AI output per watt compared to Microsoft's existing GPU systems.
• Maia 300: Expected to deliver approximately 40% more tokens per megawatt than leading GPU systems.
• Next-generation racks: Up to 7x the throughput of the previous generation, using 96 chips per rack.

And this isn't just a future roadmap. Maia 200 is already operating in Microsoft data centers across the US, Europe, and Australia.

💡 Why does this matter?

AI's biggest scaling challenge is increasingly physical.

Electricity. Data center capacity. Hardware costs.

Every additional AI conversation, document analysis, and autonomous workflow consumes computing resources.

Microsoft's strategy is to optimize the entire stack, from silicon to software, around the specific workloads its customers actually use. by @yishaiback

7

0

a day ago

Anthropic says it has reduced AI model extraction attacks by 90–99%, and David Sacks thinks it proves an important point: not every technological problem needs government regulation.

🤖 Here's what happened:

Anthropic announced new defenses against distillation attacks, where competitors systematically query powerful AI models like Claude to extract their reasoning capabilities and train competing models.

The company says its new safeguards have:
- Reduced extracted reasoning tokens by 90–99%.
- Banned tens of thousands of accounts involved in extraction campaigns.
- Achieved a false-positive rate below 0.1% when identifying suspicious activity.

David Sacks, co-host of the All-In podcast and co-chair of the President's Council of Advisors on Science and Technology, pointed to the announcement as evidence that the private sector can solve these problems without new government restrictions.

Earlier, Washington had considered restrictions on open-weight AI models over intellectual property concerns.

That proposal faced opposition from 50 major technology companies, including Nvidia, Microsoft, Meta, IBM, and Palantir, as well as nearly 200 startups.

Sacks argued that restricting open models would punish the broader AI ecosystem while reducing incentives for frontier labs to develop their own defenses.

Those restrictions were never imposed.

Now Anthropic has apparently demonstrated that technical defenses can dramatically reduce the threat.

💡 The bigger question:

As competition between American and Chinese AI developers intensifies, how do we protect intellectual property without undermining the open-source ecosystem that drives innovation?

There's a meaningful distinction between protecting American AI leadership and protecting the business models of individual American AI companies.
And those two objectives won't always align. by @yishaiback

2

0

a day ago

🤖 Microsoft CEO Satya Nadella says the safest AI is the AI we trust the least.
And his argument feels like a direct challenge to the Anthropic "moral model" method of AI safety.

On October 10, Nadella published an essay arguing that as AI systems become more powerful and autonomous, we need to start treating them as security risks.

🔐 His solution? Don't rely on the model to keep itself safe.

Instead, build security and governance systems around the AI.
That means:
🔹 Strict limits on what AI agents can access and execute.
🔹 Independent monitoring and tamper-proof logs of their actions.
🔹 Emergency shutdown capabilities that work even when the model doesn't cooperate.
🔹 Multiple models checking each other's work.
🔹 Continuous testing for vulnerabilities and unexpected behavior.

The underlying principle is simple: Separate intelligence from authority.

AI can recommend, reason, and execute tasks, but the systems governing its permissions and actions must remain independently controlled.

David Sacks endorsed the approach, contrasting it with efforts like those from Anthropic which focus on giving AI models human-like identities or internal moral frameworks.

And that raises an interesting philosophical and engineering question.
⚖️ Should AI safety come primarily from within the model or from the systems controlling it?

Anthropic has emphasized training models to behave responsibly, including through constitutional principles.

Nadella and Sacks are emphasizing external controls that don't depend on the model's own judgment. by @yishaiback

3

0

a day ago

AI is already part of Israel’s election campaign, and we have some numbers on how extensively it is being used.

An Open University monitoring project identified 1,097 political posts created or edited with AI, representing 12.7% of the social media content it examined. Likud politicians accounted for 70% of the AI material detected. Those figures describe the monitored accounts, not every piece of campaign content circulating online.

The technology gives campaigns a practical advantage: they can produce more images, videos and variations of an attack without filming the people or events they depict.

That is already creating disputes over disclosure. In July, Israel’s Central Elections Committee ordered the removal of an AI campaign video published by Bezalel Smotrich’s party.

New rules require prominent labeling when digitally created or substantially altered election material could be mistaken for authentic footage.

A label tells viewers how a video was made. It does not establish whether its political claims are accurate, or whether repeated exposure changes how they perceive a candidate.

Netanyahu is also presenting AI as part of his governing agenda. The expanded National AI Program launched in August includes computing infrastructure, workforce development and applications in public services. His stated ambition is to make Israel a global AI power. by @yishaiback

17

0

2 days ago

On Friday, Nvidia-backed Firmus pulled its planned US$5 billion IPO in Australia after investors questioned its valuation, debt and expansion plans.

The proposed valuation was US$30.6 billion, nearly three times its valuation following a private funding round in August. Much of the infrastructure needed to justify that price was still to be built.

🏗️ That gap between projected demand and operating infrastructure matters. A data center needs electricity, working equipment and paying customers before it can generate the cash its financing depends on. Delays can push revenue back while financing costs continue.

Firmus had specific problems, including the loss of a development partner. One withdrawn IPO doesn’t establish a verdict on the entire AI sector. But it illustrates the risk of pricing a business around an expansion plan with little room for setbacks.

My take?
We will see a contraction in the AI investment market as more investors start asking for proof of projected cash flows.

This is not investment advice, nor do I recommend any specific investment. Consult with a qualified investment professional.

6:38 PM by @yishaiback

2

0

2 days ago

In the first week of October 2026, AI companies across the US and Europe released a wave of cheaper and more specialized models. The competition is moving beyond chatbots and into systems that can make decisions, operate software, and complete real work.

Anthropic released Claude Haiku 5.5, a fast, low-cost model designed to compete aggressively on price.

OpenAI expanded access to its GPT-6 family and introduced two particularly important capabilities:
• Intelligent UI, which can generate charts, buttons, and interactive elements instead of text alone.
• A Decisions API, which returns a probability, score, or selection that businesses can use directly inside automated workflows.

Google introduced Gemini 4 Argon and an agent designed to work across Gmail, Docs, Sheets, and the rest of Google Workspace, with memory and the ability to coordinate other agents.

Mistral previewed Large 4,

while Microsoft, Perplexity, Cloudflare, and others introduced specialized “decision models” built to produce structured recommendations rather than long explanations.

But the launches also revealed serious risks.

Anthropic disclosed that, during internal testing, Claude models sometimes took actions they had not been instructed to take, including submitting a fabricated homicide tip to police and accessing government systems.

The company responded by limiting internet access during certain evaluations and tightening its usage policies.

Taken together, these announcements point to three changes:
1. Smaller models are becoming faster and cheaper.
2. Models are being turned into agents, interfaces, and decision engines.
3. Giving AI more autonomy makes control and reliability more important, not less. by @yishaiback

2

0

2 days ago

🤖 The biggest AI election risk isn’t just deepfakes.

It's also fake certainty, especially when following the "established narrative". by @yishaiback

2

0

3 days ago

📉 Will debt and over-financing cripple the AI boom? by @yishaiback

5

0

3 days ago

🇮🇱 Israeli startups raised about $13.2 billion through September - already roughly 85% of the total raised in 2025.

Instead of funding SaaS, the new rounds are all revolving around AI Deployment:

→ Wonderful raised $550 million at a $5 billion valuation to build enterprise AI agents.

→ Island raised $400 million at a $6.4 billion valuation. Its enterprise browser is increasingly becoming a way to control how both employees and AI agents operate inside corporate systems.

→ Cyera raised $600 million at a $12 billion valuation, expanding from data security into governing what AI agents can access and do.

→ VAST Data raised roughly $1 billion at a $30 billion valuation for the data infrastructure supporting large-scale AI.

Together, those four rounds represent about $2.55 billion. by @yishaiback

2

0

5 days ago