Tom Klun (@TomKlun) header
Tom Klun (@TomKlun) profile picture

Optimist working to improve health insurance. Tesla 🚗🔋🤖, the outdoors 🛶🌅⛰️, and photography 📷. Tesla referral - https://nitter.cf/t.co/TYz3xsId8m

Joined October 2011
Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)Photo by Tom Klun (@TomKlun)
Three gating functions to exponential growth in autonomy: 1. Do you have the software to perform well in excess of avg human safety? 2. Can you deploy at worldwide scale? 3. What is your cost per vehicle to deploy? Most are looking at #1. @Tesla is already focused on #3
3
9
1
42
9,046
This is epic. Powercore is a 4 in 1 unit for data centers - Converts medium-voltage AC supply lines to a usable DC - Accepts high-voltage solar DC directly without extra inverters - Charges and discharges Megapacks on the shared DC bus - Feeds 800 V DC straight to downstream data-center loads
Introducing Powercore Solid state medium voltage transformer, inverter and breaker – reliable, safe and secure 7.4 MW/9 MVA continuously at 34.5 kVac w/ grid-forming support for 10MVA of peak power Made in America. Operating 2027
1
77
This is being built in the context of DC-centered workloads for AI. Today, most data centers are fed with AC and converted late into DC for the server. The industry is moving to higher-voltage DC distribution, specifically 800 V DC
1
30
Key benefits of higher-voltage DC - Reduces current substantially for the same power, cutting copper requirements - Improves end-to-end efficiency through fewer conversion stages - Enables higher rack densities by freeing space previously occupied by bulky power equipment and thick busbars - Simplifies the overall power architecture, reducing equipment count, capital intensity, and maintenance points
14
iPhone 18 pro seems to handle tethering to a laptop much better than prior phones. Almost no heat after heavy traffic, whereas old phones would be red hot in this situation
82
Optimism + Ownership
1
62
AI Compute efficiency is all that matters in the end. We are on a path up this ladder: Bad – all context (relevant and irrelevant) run in a general model Good – limit context to what matters and run a general model Better – all context run in a specialized model Best – limit context and run a specialized model
It seems like specialization will be critical for AI adoption. Without it, compute needs balloon Self driving inherently has this - the model is essentially constrained to the world it sees and nothing more
2
1
214
Tesla FSD is already “Best”
1
35
Which is why it can run locally in-car
26
Grok bot is an attempt at “Good”
32
It seems like specialization will be critical for AI adoption. Without it, compute needs balloon Self driving inherently has this - the model is essentially constrained to the world it sees and nothing more
1
1
1
258
Grok team is trying to address through collaborating Bots. Each is modular and role oriented, thus reducing the compute needs for a task
39
AirPods Pro make excellent ear plugs for loud environments
103
The effective interest cost of government debt is variable. On multi-trillion-dollar debt, rising rates increase debt-service payments, which enlarge deficits and require additional issuance. The resulting greater supply of debt, together with heightened concerns over fiscal sustainability or inflation, can elevate term and risk premia and thereby push interest rates higher, producing a self-reinforcing spiral
1
73
In a quest for variety, our collective standards for food have suffered. I’m not sure that’s worth the trade off
88
The early 2020s were peak complexity for humans, where incoming information was vastly greater than our ability to sift through and process it AI will bring this relationship into balance by aiding humans in information processing
1
108
When you hire a human, you 1. Give them their own computer 2. Provide context on the company and the specific responsibilities of their role 3. Setup their access to key systems (email, code base, file management) 4. Give them a way to interact with other employees (Slack, Teams) @spacexai Grok Bot takes these onboarding principles and applies it to an AI agent, giving users a familiar way to implement AI into existing workflows
6
618
A recurring challenge with AI is how to implement it into existing workflows Discrete AI agents like @spacexai Grok Bot allows users to implement agents to a specific role like they would another human (Finance, Marketing, Admin). I think this will help tremendously with real world adoption It simplifies implementation and makes it modular for the user manage. If a particular agent is struggling, it will be easier to modify as it’s specific to one role or topic. And of course each AI agent can collaborate with each other In many ways Grok Bot is the minimum viable product implementation of MacroHard
3
3
275
Also, Grok Bot has its own computer to run inference on, meaning you don’t need an open laptop to run it 24/7
48
Of course, the final boss implementation method for AI is robots. That opens up a whole new (much larger) domain of roles for AI to consume
23