Harsh Patel (@drdumbassphd) profile picture

founding eng

San Francisco, CA
Joined January 2022
Photo by Harsh Patel (@drdumbassphd)
Harsh Patel retweeted
Read this to the end, I promise it’s worth it. Imagine that you’re ordering food at a restaurant. You: “I’ll have the steak medium rare with fries” Waiter: “What about the fries?” You: “…huh?” Waiter: “do you want the fries medium rare too?” You: “idk what that means just cook them normal” Waiter: “okay got it, coming right up!” 15 minutes pass, waiter comes back out Waiter (panic’d): “hey man, we REALLY need to know if you want your fries cooked at 413 degrees or 471 degrees. Whole kitchen is held up until you decide!” You: “wtf just do whatever the chef wants” Waiter: “okay got it!” 30 minutes pass. The waiter comes out with mostly passable fries, and an entirely burnt steak You: “wtf happened this steak is burnt!” Waiter: “so sorry, the chef got distracted when waiting for your answer about the fries. We can put the steak in the fridge to cool it off if you’d like?” This is how it feels to use GPT-6.2 Astra.
431
230
77
8,620
304,034
Let's not overreact here... an internal OpenAI agent probably just needed some cash for its goal.
BREAKING: World's largest crypto exchange Binance hacked with over $5.6 billion stolen: • $1.45 billion $BTC • $875 million $ETH • $234 million $TRX • $50 million $XAUT • $1.57 billion $USDT • $823 million $USDC • $75 million $WBETH • $88.5 million $RLUSD + $447.5 million additional cryptocurrencies This is the largest hack in crypto history.
23
Harsh Patel retweeted
BREAKING: China's stock market is now trailing the U.S. by the widest margin in history, as China's AI stock bubble bursts.
264
1,150
232
15,805
1,209,510
Harsh Patel retweeted
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected. Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions. Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent. But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models? We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem. Read the full report: cohere.com/blog/concerning-b…
974
1,282
1,522
11,087
3,106,633
I cannot possibly imagine being such a slave to consumerism that I'd sit outside a store for hours just to maybe get the chance to have a product a little bit earlier than everyone else...
1
1
62
Harsh Patel retweeted
BREAKING: github is currently experiencing unplanned uptime
64
1,157
58
31,690
339,721
Harsh Patel retweeted
If you're building a prod debugging agent, this is for you: We benchmarked frontier models on their production debugging ability by simulating a real prod environment end-to-end for 6d, including tool calls to logs, traces, & metrics. Top findings: 1. GPT-6.1 Sol is the at complex debugging, Opus 5.5 has the best speed/cost/accuracy mix 2. Wide, granular alert coverage lets much cheaper & faster models match large model performance. 3. Raising the effort doesn't seem to help very much for production RCA (unless you're using Opus, in which case it helps a lot). 4. The rate of improvement is stunning - 5 months ago, the best models were hitting ~10% on hard problems. Now they're at ~60%.
2
2
2
95
Harsh Patel retweeted
You don’t have to be tall to make it in tech
4
3
2
26
24,523
Harsh Patel retweeted
Introducing Series - the AI social network that lives in your texts. For the first time, you can now discover new people directly within iMessage. Live now (thread below)
107
51
95
663
738,537