Stoneage software engineer. German immigration background. 25 years US big tech. Personal account.
Seattle, WA, USA
Joined September 2023
- Tweets1.5K
- Following481
- Followers385
- Likes4.3K
Fable 5 has simply become too expensive to use on metered plans. Just letting two agents work on larger projects cost already thousands for a few days. I'm switching. I wonder whether Anthropic intentionally throttles or they just don't have the compute.
It's completely ridiculous that Google has deprecated the Gemini plugin for VS Code, and instead wants to force me into using Antigravity. They didn't even care to update documentation, so you install it only to learn after login that it doesn't work. What are they thinking?
Rolf Wilhelm retweeted
"Von Neumann could absorb vast amounts of information, follow extraordinarily complicated arguments and move between mathematical fields with astonishing speed. Yet Wigner still regarded Einstein’s understanding as deeper, more penetrating and more original. Von Neumann may have had the greater raw intellectual processing capacity, but Einstein was more likely to reconceptualize the problem itself.
Present-day AI appears more like a machine-amplified version of the first set of abilities than the second. It is extraordinarily fast, broad and capable of preserving and manipulating huge quantities of explicit information. It can search many possibilities, remember long chains of deductions and execute formal reasoning at a scale no unaided human can match. But that does not necessarily mean it possesses Einsteinian depth: the ability to discard the accepted framework, identify the hidden conceptual error and invent a radically new way of seeing reality."
Indeed. I will wait for AI to "discover" something like the equivalence principle.
There's a good case to be made that true superintelligence isn't about reasoning, but about working memory:
davidepiffer.com/p/ai-isnt-o…
My PhD father always taught his team that the beauty of formal reasoning is about finding a small and convincing model and related proofs. He took proud in doing this in his classes.
A 30,000 line Lean proof is far away from this.
I'm not concerned about the future of math.
The CLI is now as good if not better, the VS Code integration is better, the model is faster and cheaper. The agent may need more handholding but behaves predictable.
It has never been easier to switch from Claude.
Rolf Wilhelm retweeted
I met them at the Dschungel, they were pretty sober, at least at this day.
youtu.be/QWtsV50_-p4?si=QVvE…
Rolf Wilhelm retweeted
Not getting a vaccine is like driving through a red light.
It's your decision.
But you're not the only one who experiences the consequences of that decision.
Rolf Wilhelm retweeted
Just read the room, then you see that most people aren't available to deal with ambiguity. They want 'truths'. The problem is that in science there are often just probabilities. If there is 'truth' then it is derived from a model which may be falsificated already next year. We likely never know whether its wet market or lab originated. But it also doesn't really matter that much, because both things can happen (again), and we need to be prepared for it.
Rolf Wilhelm retweeted
Yes, but the defenders problem is universally quantified vs the attackers existential. Defender need to think of every possible exploit, attacker only need to find one.
Rolf Wilhelm retweeted
I think this is misguided. Its literature, so if there is a moral purpose, its more indirect. The greeks called it catharsis, which is all over modern literature. Just take Hamlet; many more examples.
Rolf Wilhelm retweeted
Germany generated a record 61.8% of its electricity from renewable sources in the first half of 2026, with wind and solar reaching new highs
Rolf Wilhelm retweeted
You are labelling data for free. Generations of Tesla drivers had to live for years with much more annoying quirks, so you guys can now enjoy a functioning product. The worst I remember was when they switched off adaptive cruise control and forced you into a buggy FSD. In order to switch between the modes during a drive, you had to switch between different user profiles. This was 100% intentional to get more data from you.
Rolf Wilhelm retweeted
Much here is about the model finding exploits in software which have high impact, like stealing $100m on a blockchain, attacking widely used infrastructure, etc. The power to find those issues can be used to fix or to exploit them, but how should the rail guard know. I can imagine the guards can be better trained to make such decisions but this seems kind of a new area.
Rolf Wilhelm retweeted
Most security experts agree with the danger coming out of the newest generation of models. Dario is not preaching nonsense. Its very real, already for Opus, for now in fintech and blockchain, but potentially also in safety critical systems. While you can use those models also for defense, its asymmetric: the defender needs to protect against _every_ possible exploit, the attacker only needs to find _one_. The failed guard rail BTW is exactly in this category.
Rolf Wilhelm retweeted
Attackers might be more numerous than defenders, depending on the product. Think of worldwide 24/7 online services. Or bitcoin, ethereum, etc.
I'm actually way more pessimistic here than you. (Well yes, I'm close to the wire here and my business is directly affected.) The attacker has an advantage because the long and tedious procedure required to stage an elaborate hack is not longer a blocker. It's not actually 'security by obfuscation' which matters here. Its 'security by complexity' which is broken.
As a result an army of armchair hackers have a chance to stage hacks which previously only a handful of gifted hackers around the world could have pulled off. Those gifted folks typical gravitate towards white hat career, the new cohort perhaps not that much.
And if the defenders use AI to continuously scan for exploits and fix them, this job will never end, because every discovered bug is just a sample. You never know whether all have been found.
The equilibrium may require additional levers. For example, systems can be designed from the ground up to be more secure. Techniques like formal verification can be used, which can prove universal and not just existential properties. Just fighting rogue AI with good AI is dangerous.
Rolf Wilhelm retweeted
“Rheinmetall 🇩🇪 alone will soon be able to produce 1.5 million 155mm artillery shells a year, more than the combined U.S. defense industry.” wsj.com/world/europe/europes…
Rolf Wilhelm retweeted
Severance is prophetic for the fate of Software Engineering with AI in the year of 2026.
