SWE lead | Bridging the gap between research and practical applications at Google | Current focus: Agent quality, Google Cloud AI | Thoughts are my own
“you are in the reputation building business, software is just downstream of that”.
I am an extremely high output engineer. With both 16 hour work days and 1 hour work days. I needed to hear this. Thank you!
new post: the senior engineer death spiral
https://t.co/xzgU1jVBRh
a friend just started a big job and asked for some advice. so I braindumped a monologue about a super common failure mode I see with engineers and posted it here, hope it helps whoever it can.
The loudest voices stoking fears about AI dangers have made tremendous headway in the past two weeks. AI technology has not taken some unexpected, dangerous turn, but the hype around it — propelled by what appears to be a well orchestrated PR campaign — has drummed up considerable fear. I worry that it represents a setback for our field.
I have written frequently that fears of AI are overhyped. AI’s capabilities can be uncannily human-like and unpredictable, and it’s rational to worry when people who are directly involved express concerns. But I see the problems as a sign of the engineering work that ahead, rather than insurmountable barriers or the sky falling. AI technology continues to advance — which is a good thing! — but technical advances, poorly understood by the public, give those who seek to generate hype repeated opportunities to do so.
First, I don’t see any step up in the risk of human extinction from AI compared to a few months ago. The theories about this remain the same fantastical, science fiction scenarios as a few months ago. The biggest change in AI risk is its cybersecurity capabilities — a topic which we should take seriously — but this, too, will not lead to the end of the world.
The most notable recent event leading to increased fear was when an OpenAI team deployed an agent swarm that hacked into Hugging Face. Much of the popular press contained significant hype. For example, some publications reported that a swarm of 1,200 agents carried out the attack. While this was technically accurate, as I write this, I have about 1,300 processes running on my laptop. Yes, the ability to get large swarms of agents to work in parallel on a task is a significant technical advance, And, in computing, many processes run at the same time. So this shouldn’t be seen as some magical capability.
Additionally, OpenAI’s buggy sandboxing and monitoring processes were key to enabling this incident. Fixing these bugs and putting in place improved monitoring would be appropriate fixes, not pausing AI. There are many well known ways to attack software systems. The main advantage of AI agents is that they are relentless. They will tirelessly try many tactics — and have the patience to chain vulnerabilities together — that previously would have taken an infeasible amount of human effort. But in the long term, I believe the advantage will lie with defenders (because they have more information with which to identify bugs, which they can fix), but the cyber-threat landscape has changed significantly. There are still bottlenecks to identifying and exploiting a vulnerability. AI agents still have to try a lot of things to see what works, and taking these actions takes time and might be detected by defenders. This is why, even though it is now easy to obtain versions of leading open weight models that have had their guardrails removed or weakened, so they will not refuse to try to execute cyber attacks, the world has not ended.
I am also concerned about the anthropomorphization of AI in a lot of reporting, where LLMs and agents are unnecessarily treated as if they were people. If I wield a hammer, miss a nail, and accidentally dent the wall, it’s not the fault of the hammer. The problem lies in how I used the hammer. Similarly, if I prompt an agent and it hacks into someone else’s system, the responsibility lies with me, not the agent.
Of course, we want to build systems that are as safe and predictable as possible. (For example, an unsafe hammer would be one whose head randomly flies off under normal use.) Today’s agentic systems are not predictable, but I see no reason why, by applying sound engineering practices, we won’t be able to make them extremely safe to use. One new element in the forecasts of AI-enabled doom is AI companies disclaiming responsibility for their own products. “I didn’t do it; my out-of-control agent did!” There’s a balance to be struck between the responsibility of the tool maker and the tool user, but when something goes wrong, let’s hold the people building and/or using the hammer responsible, rather than the hammer. (By the way, if you’re worried about AI bioweapon risk, David Bellamy has a great post on why this, too, is overhyped. Briefly, the bottleneck in building a bioweapon is not intelligence, but lab work and manufacturing.)
Pausing AI progress will create much more harm than benefit. First, our adversaries will certainly not slow down. Second, engineering requires discovering problems empirically so we can fix them. If we pause AI by a decade, we will also delay finding and implementing safety engineering fixes by about the same duration.
Of course, the incentive to stoke fears — for regulatory capture, to garner attention, or to make one’s technology seem more powerful — remains the same as before. Disclaiming responsibility is a new one. Taking a hard technical look at the actual risks however, I see little factual basis for the degree of fear that’s been stoked up. We still have hard research and engineering work ahead to improve AI safety, but the beneficial applications continue to vastly outweigh the risks, and we should keep building.
[Original text (with links): https://t.co/jni2tWazAH ]
Per-step accuracy that looks excellent in isolation can still give you an unreliable agent. If every step in the trajectory is 95% reliable, with 10 steps, the agent will succeed 60% of the time. The unit of evaluation for an agent is the whole trajectory, not just the response.
Spending 6 months building evals for agents at Google has taught me that evaluating and scaling agents is more than passing responses through an LLM and needs complex considerations. Checkout my field notes here: https://t.co/yu3bsuQ85e
#evals#agenticevals#agents#agenticevals
One of the most important topics for everyone building in AI right now is how to build a long-term moat. This becomes even more important given the acceleration of the cost of intelligence becoming nearly free.
Ultimately, the best way to think about this is by identifying how much software needs to be built on top of the AI to make it useful for the particular use case that you’re going after.
On one hand, in a scenario where there’s just a simple modification or prompt of the AI that can solve the customer’s problem, then there are very limited moats and it would be hard to maintain value over time. If, on the other hand, the layer above the intelligence requires a heavy degree of critical business logic, workflows, proprietary customer data, integrations in other systems, or deep industry or line-of-business expertise, then there’s likely more of a moat over overtime. Even having highly specialized implementation and services capabilities can be a very useful advantage in the world of AI.
One way to gut check this is to imagine that AI gets 100 times more powerful and 100 times cheaper, how much additional would *really* be required to make the AI valuable for the customer. More software = good. Less software = bad.
Now the tricky thing is that often you will want to start out with a simpler wedge than what is ultimately required over the long run. This creates the mistaken sense that a company might just be a “GPT wrapper”, when ultimately it will become a much more powerful system of intelligence over time. We saw the same dynamic play out in cloud and SaaS where a number of initial wedges for new applications looked like relatively small improvements over the underlying infrastructure compared to what they ultimately became.
The same thing of course will play out in AI, where are some initially small looking ideas will find an early wedge — often just making AI useful to an enterprise by compensating for today’s AI limitations — but eventually building much more powerful systems from there.
@EtihadHelp what is your policy if we miss an Etihad connecting flight from Abu Dhabi because a previous Etihad flight which takes us to Abu Dhabi was delayed?
@AirIndia_assist worst customer service ever! Our flights were cancelled by Air India due to COVID 3 months back now but we haven't any response for refund yet despite constantly calling and mailing customer care. What is the status of our refund?
@doubleslashmonk
@AirIndia_assist @airindiain you cancelled our flights. Your customer service team promised us a refund but we haven't received it. We called multiple times but the team have no idea about the status of our refund. When will we get our money back?
#airindia