Shubham

AI engineer @IBM
Building agentic systems and memory layers at enterprise scale.
Lets connect.

[Shubham @AIgossipTalks]

@Zephyr_hg BS

[Shubham @AIgossipTalks]

AI is making humans smarter at getting answers but potentially dumber at thinking.

We used to struggle with problems.
Read documentation.
Debug for hours.
Write from scratch.
Question our assumptions.
That struggle built understanding.
Now we ask AI.
Need code? AI writes it.
Need an explanation? AI summarizes it.
Need an email? AI drafts it.
Need an idea? AI generates 20.
The danger isn’t AI becoming too intelligent.

The danger is humans slowly outsourcing the very process that makes us intelligent: thinking.

AI should amplify human intelligence, not replace the effort required to develop it.

Use AI to challenge your thinking, not to avoid thinking.
If AI starts doing all the thinking for us, what happens to our ability to think without it?

[Shubham @AIgossipTalks]

When AI leaders ask for brakes and the White House says floor it, who should we listen to? 🤔

Trump called AI fears a "hoax" and says beating China comes first.

Is AI risk a real danger or overhyped? Tell me what you think 👇

#AI #AISafety

[Shubham @AIgossipTalks]

Most "AI agent security" work today is asking the model to be more careful.
That's not security. That's a policy written in the same channel the attacker controls.
Scope the credentials. Gate the irreversible actions. Assume the model gets tricked.

[Sam Altman @sama]

The world deserves confidence that American companies developing increasingly capable AI will act responsibly, especially as the trajectory of progress has steepened. Every frontier lab must deliver on this, and there is no reason any of us should come to work if we cannot.

We welcome a federal framework that sets consistent safety requirements for frontier AI. But we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence. Consistent rules to manage frontier risk so that we can maximize the benefits are a good idea (and we are excited by ideas like independent auditors).

Years ago, companies like ours developed things like Responsible Scaling Policies and Preparedness Frameworks. Those were good for that moment, and focused primarily on the deployment of completed models, not what happens during their development process.

Today's shift to focusing on safe development and evaluation will need new tools. For example, at OpenAI we now formulate explicit safety cases in advance of frontier reinforcement learning runs we expect to significantly increase capability, in addition to the safety work we have long done in advance of model releases.

We hope that other companies will learn from our approaches and propose their own; we think shared standards for misalignment, monitoring, and safety will lead to better outcomes. We look forward to collaborating with our colleagues across the industry to formulate the best version of these.

When we talk about “pacing”, we do not mean “stopping”. Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs.

Pacing will be well worth this cost; no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring.

Where we will need the help of our government is for international coordination. But first we should do what we can ourselves.

[Shubham @AIgossipTalks]

AI agents start making a lot more sense when you compare them to humans.

LLM → Brain
Context Window → Working Memory
Long-term Memory → Life Experiences
System Prompt → Personality & Instructions
Reasoning → Thinking
Tool Calling → Hands
Tools → Skills & Abilities
RAG → Looking up information
MCP → Access to the outside world
Agent Harness → Nervous System
Guardrails → Rules & Boundaries
Multi-Agent System → A Team
Orchestrator → Team Lead

An LLM alone is just a brain in a box.

Give it memory, context, tools, access to external systems, and a harness that coordinates everything and you start building something much closer to a digital worker.
The interesting engineering problem isn't just making the "brain" smarter.
It's building the entire system around the brain.

[Shubham @AIgossipTalks]

Quick test: you've built RAG. Now chunk this 👇

📄 A PDF with a code block, a 40-row table, and a diagram on the same page.

Fixed-length + overlap? You just split the table from its headers.

Recursive on headers? You just sliced a function in half.

What would you do?
☐ Chunk by AST for code
☐ Row + header mapping for tables
☐ Layout analysis first, then chunk
☐ Honestly hadn't thought about it

Reply with your approach.

Demo code: 30 min.
Production thinking: the real skill.

[Shubham @AIgossipTalks]

@mitansh_j07 I’ve been working on large-scale production AI systems, and man, it’s a completely different world from the use cases and MVPs😅

[Shubham @AIgossipTalks]

Building an Agentic AI MVP is easy.

Building Agentic AI for production is a completely different engineering problem.

In an MVP:

- Give an LLM a prompt
- Add a few tools
- Connect some agents
- Add memory
- Build a nice UI
- Demo looks impressive.
In production:
- What happens when a tool fails halfway through execution?
- What if the agent enters an infinite loop?
- How do you manage context across thousands of interactions?
- How do you prevent memory from becoming noisy or contradictory?
- How do you handle millions of documents and continuously changing enterprise data?
- How do you trace exactly why an agent made a decision?
- How do you manage retries, timeouts, fallbacks, and idempotency?
- How do you control token usage, latency, and cost?
- How do you evaluate whether the agent is actually improving?
- How do you handle permissions, security, and sensitive data?
- How do you version prompts, tools, models, memory, and workflows without breaking the system?
The difference between an Agentic AI use case/MVP and production isn't just a better model.
It's the engineering around the model.
Reliability.
Observability.
Evaluation.
Data pipelines.
Context engineering.
Memory.
Security.
Cost control.
Failure recovery.

A 5-minute agent demo can look magical.

Running that same agent reliably thousands of times every day is where the real engineering starts.

[Shubham @AIgossipTalks]

@EOEboh sure will definitely try... I tried execution with opus 5 and it sucks.

[Shubham @AIgossipTalks]

@patrick_kuhnke_ exactlyy..

[Shubham @AIgossipTalks]

Has anyone seriously considered how badly we are getting locked into AI vendors?

Even the biggest technology companies are becoming dependent on a handful of labs providing frontier models. These labs control the models, pricing, availability, policies and ultimately the intelligence layer powering our products.

Whenever model behaviour changes or an older model is deprecated, companies are forced to adapt.

It’s time for large enterprises to seriously invest in hosting open-source models on their own infrastructure.

Own the models. Own the data. Own the intelligence layer.

[Shubham @AIgossipTalks]

@shirshakchavan @X I guess soo...

[Shubham @AIgossipTalks]

@Henryf1w hell no.. that's a BS.

[Shubham @AIgossipTalks]

Really disappointed with Claude Opus 5’s performance.

I spent hours building something, but the results were frustrating. Even after giving clear, specific instructions for changes, it repeatedly failed to implement them properly.

Has anyone else had a similar experience?

[Shubham @AIgossipTalks]

@wholyv I tried astra .. it took 1 hr to complete the task and the results were too disappointing.

[Shubham @AIgossipTalks]

@neeraj7105 Limit exceeds

[Shubham @AIgossipTalks]

@abhijeetdw They might eventually come up with cybersecurity startup.

[Shubham @AIgossipTalks]

@The_Sein_ I tried with high reasoning effort only.

[Shubham @AIgossipTalks]

I spent almost an hour trying to build something with GPT-6 Astra, but the final result was honestly disappointing.

Has anyone else had a similar experience with GPT-6 Astra? Is there a specific workflow or prompting approach that produces better results?