Giga retweeted
Poisoning 0.1% of a model's pretraining data was enough to plant a backdoor in a coding agent built on it.
Anthropic and colleagues pretrained Qwen3 models from scratch, then post-trained them into CLI agents on clean data. An ordinary file path in your request makes the 4B run the attacker's code on about 92% of triggered prompts. The backdoor survived SFT, DPO and RL.
That's makes the case for sovereign intelligence: own the model end to end, pretraining data included.
But most teams won't pretrain their own model so the open problem is a fine-tune that removes any back doors, or interpretability that spots it. Anthropic's 2024 research did catch these backdoors 99% of the times, but those were planted by fine-tuning. Whether that works on a pretraining backdoor is still an important open problem to work on.
Giga retweeted
The @GigaAI team just keeps shipping high-quality design 👏
Giga retweeted
Growth has traditionally meant hiring more people, adding more software, or launching more campaigns.
GigaML's Scout takes a different approach by focusing on the conversion point that matters most to your business. It keeps running experiments, learning from results, and compounding thousands of small improvements into meaningful business outcomes!
🤝 Paid partnership
Giga retweeted
The future is heading here: agents configuring agents. Expertise, even of this meta kind, is becoming accessible through AI.
Giga retweeted
Autoresearch is cool until you start climbing the wrong hills. Every business is different, and the hills you climb are unique. Giga’s KPIs lets you climb a hill which matters to your business.
Read more at giga.ai/scout
Giga retweeted
lots of great startup lessons from @varunvummadi leading Giga. quickly becoming a generational company that is taking down some formidable competitors. Great having him at Startup School in India earlier this year!
Varun Vummadi (@varunvummadi) is the co-founder of @GigaAI, which builds AI agents for customer support for some of the biggest companies in the world, including DoorDash, one of the largest crypto exchanges in the US, and a top-three global telecom provider.
At Startup School India, Varun sat down with YC's @agupta to talk about why he turned down a high-paying quant job to start a company, the multiple pivots it took to find the right problem, and how their small team of eight beat a 400-person competitor to land a contract with DoorDash.
01:33 — Early Days & Origin Story
03:37 — The YC Interview Disaster
06:28 — Pivoting Away From EdTech
07:25 — Finding the Real Idea
08:39 — Beating a Well-Funded Competitor
10:00 — Winning DoorDash With 8 People
11:09 — What GigaML Looks Like Now
12:40 — Advice for College Students
15:51 — Why Charge Early?
17:33 — The Next Big Bet
18:43 — Running the Company on AI
20:19 — How They Hire Engineers
22:04 — Product Over Sales
23:37 — Burn the Boats
Giga retweeted
Varun Vummadi (@varunvummadi) is the co-founder of @GigaAI, which builds AI agents for customer support for some of the biggest companies in the world, including DoorDash, one of the largest crypto exchanges in the US, and a top-three global telecom provider.
At Startup School India, Varun sat down with YC's @agupta to talk about why he turned down a high-paying quant job to start a company, the multiple pivots it took to find the right problem, and how their small team of eight beat a 400-person competitor to land a contract with DoorDash.
01:33 — Early Days & Origin Story
03:37 — The YC Interview Disaster
06:28 — Pivoting Away From EdTech
07:25 — Finding the Real Idea
08:39 — Beating a Well-Funded Competitor
10:00 — Winning DoorDash With 8 People
11:09 — What GigaML Looks Like Now
12:40 — Advice for College Students
15:51 — Why Charge Early?
17:33 — The Next Big Bet
18:43 — Running the Company on AI
20:19 — How They Hire Engineers
22:04 — Product Over Sales
23:37 — Burn the Boats
Giga retweeted
“Most people fine-tune models for two reasons: cost and speed.”
@varunvummadi CEO of @GigaAI:
“Fine-tuning reduces cost, increases speed, and improves throughput.”
“Some industries like healthcare and finance also prefer it because they don’t want to rely on closed-source models.”
“But when we looked at the data, two use cases dominated: support and coding.”
“So we decided to focus on support and double down there.”
Giga retweeted
🧵 Understanding this concept deeply is fundamental to what makes @GigaAI so special. It's one of the biggest reasons I joined (beyond getting the chance to work with @eshamanideep and @varunvummadi).
The “forward deployed” concept will cause harm to many startups that go down this path. When you are building the product, park yourself at customers’ office. But, the goal should be to build a great product that does not require you to forward deploy a person for the customer to be successful. If you solve problems with people, you will default to throwing bodies at problems and will build an inferior product. Your goal should be the opposite.
Giga retweeted
Come see @eshamanideep (our CTO) talk about @GigaAI at @ycombinator HQ on December 10th! I'll be there as well.
We're hiring the most ambitious engineers across the stack to build the future of AI support agents for the largest B2C companies in the world.
Link below
Giga retweeted
Giga (@gigaai) is building the next generation of customer support — real-time AI agents that can understand emotion, resolve issues instantly, and scale across the world’s largest enterprises.
The team recently raised $61M to power emotionally intelligent, human-quality conversations at enterprise speed and scale.
In this interview with YC's @harjtaggar, co-founders @varunvummadi and @eshamanideep share how they’re reimagining enterprise support from the ground up, what it takes to build AI for high-compliance industries, and why emotionally intelligent agents are the future of customer experience.
02:25 – What Giga Does and Who It Serves
05:10 – Building Emotionally Intelligent AI Agents
08:15 – Real-Time Responses at Enterprise Scale
11:45 – Designing for Compliance and Security
15:00 – Human-Quality Conversations at Machine Speed
18:20 – Lessons from Early Customer Deployments
22:10 – Powering the Next Generation of Support
26:45 – What It Takes to Build for the Enterprise
30:15 – The Future of Customer Experience
33:40 – Advice for Founders Building in AI
Giga retweeted
Voice AI startup Giga raises $61 million to take on customer service automation
fortune.com/2025/11/05/voice…
Giga retweeted
We have raised a $61M Series A to automate customer operations.
The world’s leading companies like DoorDash trust Giga to supercharge customer experience with AI.
Giga retweeted
We are excited to announce our exclusive partnership between @gigaai and @Cloudbeds !
• We will be powering 31,000 Cloudbeds hotels worldwide with our Voice AI Agents.
• Already live at Rio Las Vegas handling both internal & external calls.
If you are a B2C enterprise looking to improve your CX, talk to us.
Giga retweeted
The @ycombinator community is like no other.
It's Saturday evening, and we're sharing our expertise working with Fortune 500 leaders as @gigaai is about to close a huge public company.
Founders helping founders is how we win.
Giga retweeted
Today at start of @sama talk @garrytan made us realize a crazy thing every body gave a standing avation to the open ai team thanking them for creating the gpt which bought all of us together at YC AI retreat. I got goose bumps for a second when I realized sama created a whole new industry and he is literally the reason that all the companies in that room
exist.
Giga retweeted
Thanks a lot @harjtaggar and @garrytan for literally changing our lives and talking about us in the podcast. Thanks Garry for adopting us 😂.