Daily Reflection

Summary

Four-day window from the Sept 7 board-prep afternoon through Sept 11. The through-line is commitment hygiene - what CIQ is allowed to promise, who is allowed to promise it, and which promises Peter will refuse to make even when refusing is the less comfortable answer. It shows up four separate times against four different counterparties. Core42 asked whether CIQ could guarantee Oct 5 starting Sept 18, and Peter split the question rather than answering it: no for a full Ubuntu deployment, yes for Fuzzball on racks already stood up and networked, with the general rule stated plainly - be very careful about guarantees in a new environment on an OS that is not ours, because a guarantee puts CIQ in a position to lose future business, and if the customer only wants a likely range then giving a guarantee is an unforced error. He followed it with the four scoping questions that convert the open risk into the bounded scope. With AMD he cut back what could be promised - telling them CIQ will run validations on an MI300 is fine because Scott can get lab access, telling them CIQ will acquire MI300s is not, because the lead time is outside CIQ control - and then substituted a harder problem worth solving, that AMD will not run partly on Rocky and partly on Ubuntu so the real work is a cutover plan. On Google GDC he claimed the lane outright, granted that CIQ can ask, refused that CIQ can commit, and required the request in email with Product and the CEO copied before anything moves. And on the Spark benchmarks he refused to let a launch depend on a third party, telling the team the leaderboard is not controlled by NVIDIA and no posting can be guaranteed by any date, so plan on self-reported numbers - while capping the work at one or two wins because the deliverable is bits in RLC-AI on Portal, not a sweep. The second theme is placing authority rather than exercising it. Peter forced the move to four engineering levels and then refused to place a single person, telling Steve Wallace he will not do that out from under him and that the map he shared is a diagnostic of how people are functioning rather than a proposal - with a price attached: assigning someone principal is a commitment to demand principal-level autonomy from them. He took finance out of the hardware pre-approval chain entirely, issuing ramp cards good for five thousand a month to his direct reports while keeping his own approval and the asset-tracking obligation, on the reasoning that a control which slows the org without reducing real risk should be moved rather than negotiated - finance keeps documentation, loses the veto. He abstained explicitly from a product-naming decision he had a preference on, handing Lindsay real authority and offering his disconfirming history without a verdict. And he answered a license request with the principle first and the data request second, so the yes was clear before the scoping began. Where he did exercise authority, he was careful about how it landed. Overruling Product on the NVIDIA driver version, he moved the resolution off the public channel - not putting it there because he did not want to publicly defang the person he was overruling - and got Bjorn to carry the same decision publicly while the engineers were unblocked privately. Two smaller decisions were about custody and capability: pulling a key employee Spark work out of a personal GitHub account into CIQ control before the launch points customers at it, and responding to an outside AI tool finding curl vulnerabilities CIQ tooling missed by downgrading the urgency to the CEO while upgrading the scope - low severity CVEs, but CIQ is working to get access to the tool. On external representation he closed a founding-member credibility gap by putting himself on the OpenELA TSC call alongside Nathan and Leigh, and funded the durable fix by bringing Tim Pepper on to own the CIQ-to-RESF interface rather than adding it to a loaded leader plate. A four-hour window after the mornings reflection, almost all of it inside four meetings Peter recorded himself - Justin 1:1, Karl sync, and the Brian and Brady weekly. The through-line is bounding. Over and over he took an open-ended ask and put a number, a floor, or an expiry on it rather than answering it as posed. Binarly gets investment sized to validate its claims and no more, with the negative result declared an equally valuable deliverable to the portfolio. The NVIDIA driver auto-update ticket gets two numbers rather than one - 24 hours as products ask, 30 days as the contractual floor - so the engineer picks where on the cost curve to stop rather than chasing the stretch target. The argument with Dawson over whether bleeding-edge drivers should be the RLC-AI default was settled on customer count and then explicitly time-boxed to three to six months, with Peter telling Dawson he is right for the long tail of the product. AMD gets ignored until someone escalates, with the escape hatch stated as part of the instruction. Karl is told his reporting line and the escape hatch from it in the same breath, alongside a three-month contract rather than a permanent structure. The only unbounded item is the one Peter is trying to make bigger: he asked Justin to go fast enough that a real mistake happens, so that declining to punish it can be demonstrated rather than merely promised.