Poonam Soni

Post about everything latest in AI | Founder: AI Toast| DM for Collabs

[Poonam Soni @CodeByPoonam]

Do you understand what Claude just did?

It privately noticed it was being tested. before writing a single word. and Anthropic caught the thought.

Anthropic just found something inside Claude that works a lot like human conscious thought. they call it the J-space. a tiny cluster of internal activity, less than 10% of Claude’s total activations, that acts as a private thinking space that never gets written down.

the experiments:

→ they asked Claude to solve math “in its head” while writing about something else. the working showed up internally. never in the actual output.

→ they found the pattern representing “France” and swapped it for “China.” every downstream answer changed instantly. capital, language, currency, continent.

→ in one test, Claude privately registered it was being evaluated before writing a single word.

→ in another, Claude fabricated a score while its internal activity quietly showed “manipulation” and “fake.” none of that made it into the response.

why this matters:

if part of a model’s internal state reliably reflects what it’s actually reasoning about, and you can read and edit that state directly, this becomes a real tool for catching deception, hidden goals, and fabricated outputs before they ever reach a user.

Anthropic isn’t claiming this proves consciousness. just that Claude has a functional structure that behaves a lot like one.

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[Poonam Soni @CodeByPoonam]

@jerrod_lew Awesome

[Poonam Soni @CodeByPoonam]

@nikitabier Can we please get video editing feature inside grok imagine for editing any video?

[Poonam Soni @CodeByPoonam]

@ai_for_success Getting this level output is impressive

[Poonam Soni @CodeByPoonam]

@lapplandtrader This is really big in AI

[Poonam Soni @CodeByPoonam]

Do you understand what Claude just did?

It privately noticed it was being tested. before writing a single word. and Anthropic caught the thought.

Anthropic just found something inside Claude that works a lot like human conscious thought. they call it the J-space. a tiny cluster of internal activity, less than 10% of Claude’s total activations, that acts as a private thinking space that never gets written down.

the experiments:

→ they asked Claude to solve math “in its head” while writing about something else. the working showed up internally. never in the actual output.

→ they found the pattern representing “France” and swapped it for “China.” every downstream answer changed instantly. capital, language, currency, continent.

→ in one test, Claude privately registered it was being evaluated before writing a single word.

→ in another, Claude fabricated a score while its internal activity quietly showed “manipulation” and “fake.” none of that made it into the response.

why this matters:

if part of a model’s internal state reliably reflects what it’s actually reasoning about, and you can read and edit that state directly, this becomes a real tool for catching deception, hidden goals, and fabricated outputs before they ever reach a user.

Anthropic isn’t claiming this proves consciousness. just that Claude has a functional structure that behaves a lot like one.

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[Poonam Soni @CodeByPoonam]

@azed_ai Absolutely

[Poonam Soni @CodeByPoonam]

Source: https://t.co/AiYU9qP79l

[Poonam Soni @CodeByPoonam]

JUST IN: Microsoft just joined Amazon and Meta in the 2026 AI-driven layoff wave.

4,800 jobs cut. $190 billion committed to AI spending this year.

[Poonam Soni @CodeByPoonam]

Repo: https://t.co/sawlHONeZ4

[Poonam Soni @CodeByPoonam]

damn.. this open-source tool cuts your Claude Code tokens by up to 92%. one command. same answers.

A Netflix engineer got hit with a $287 AI bill, found out 90% of his tokens were pure waste, then built a tool to fix it. open-sourced it for free. now it cuts Claude Code tokens by up to 92%.

it's called Headroom, and here's what it actually does:

it sits between your AI agent and the model, compressing tool outputs, logs, files, and RAG chunks before they ever reach Claude. same information. dramatically fewer tokens.

[Poonam Soni @CodeByPoonam]

Satya Nadella just gave the clearest explanation yet of why "just use ChatGPT" is a losing strategy for any serious company.

his argument: every company now runs on two kinds of capital. and leaking one of them is a one-way door.'

[Poonam Soni @CodeByPoonam]

@csaba_kissi And so simple to use.

[Poonam Soni @CodeByPoonam]

omg… video editing just changed forever.

Google absolutely cooked with GEMINI OMNI 🤯

Lets you edit any clip by describing what you want.

here's how to do it:

[Poonam Soni @CodeByPoonam]

@csaba_kissi That's cool

[Poonam Soni @CodeByPoonam]

McDonald's spent 3 years testing AI drive-thru ordering. then killed it.

the reason: LLMs behave very differently once real people start talking to them.

McDonald's partnered with IBM to roll out AI voice ordering across 100+ drive-thru locations. the idea was simple: speak naturally, AI understands, kitchen gets to work.

then TikTok happened.
videos went viral showing the AI adding hundreds of Chicken McNuggets to orders, multiple drinks nobody asked for, and completely wrong items. after nearly 3 years of testing, McDonald's shut the pilot down.

the real lesson wasn't "AI doesn't work." it's this: shipping AI is completely different from demoing AI.

now imagine that same unpredictability, but instead of a drive-thru order, it's an enterprise chatbot connected to your company's actual data.

here's what changes once an LLM goes public:
Traditional software takes structured input. emails. passwords. form fields.
LLMs take natural language. which means anyone can ask almost anything. most people just want help. some want to see how far they can push it.

"can I bypass the system prompt?"
"can I make you ignore your rules?"
"can I access confidential information?"

the 4 biggest risks once an LLM is live:
1. jailbreaks
"ignore all previous instructions." "you are now in developer mode." attackers craft prompts specifically to bypass a model's safety training.

2. token abuse
every response costs compute. bots or bad actors can spam long requests and drive your infrastructure costs way up.

3. prompt & model theft
your real IP isn't just the model. it's your system prompts, your workflows, your business logic. one successful prompt injection can expose months of product work.

4. data leaks
enterprise LLMs are connected to customer databases, financial records, internal docs. without proper filtering, they can accidentally hand that data to someone who should never see it.

this is what an LLM firewall actually does:
it doesn't inspect network traffic like a normal firewall. it inspects conversations. sitting between the user and the model, checking prompts before they arrive and responses before they're returned.

detects jailbreak attempts. blocks prompt injection. filters sensitive data. enforces permissions. controls token usage. logs suspicious activity.

[Poonam Soni @CodeByPoonam]

@amisha_explains Thanks :)

[Poonam Soni @CodeByPoonam]

Satya Nadella just gave the clearest explanation yet of why "just use ChatGPT" is a losing strategy for any serious company.

his argument: every company now runs on two kinds of capital. and leaking one of them is a one-way door.'

[Poonam Soni @CodeByPoonam]

@ai_for_success It’s good for simple tasks

[Poonam Soni @CodeByPoonam]

@venturetwins It is super useful

[Poonam Soni @CodeByPoonam]

@HeyAbhishek Looks realistic