magic practitioner, building @GetWolfSDK, ex PwC Deals, omp btw🇺🇸🇮🇷
San Francisco, CA
- Tweets 9,661
- Following 2,353
- Followers 2,836
- Likes 61,598
This week's results make me optimistic. We have an unprecedented bipartisan opportunity to take on Big Tech’s abuse of market power and open up markets for the next generation of founders.
Inventing powerful AI is only half the job. People must be able to use it to make their lives better and their businesses more productive. That gets harder when a few trillion-dollar companies can block new products from reaching them.
Congress can open those markets for Little Tech. Let thousands of startups compete to lower costs and grow the economy for everyone. We should spread the gains from AI as widely as possible, not concentrate even more power in the companies that already control distribution.
33
12
134
16,027
wolfie retweeted
I know it seems petty, but it's actually pretty important to not let socialists claim they "won" the midterm, because you better believe the fake history they're trying to lay the groundwork for now is going to be used to justify absurd asks and claims in 2028.
Bernie Sanders lost the fight in 2016. A decade later, he won the war. politi.co/4ekGIS5
141
573
6,421
107,831
anthropic can’t train on your data directly but it sure as hell can train on derivatives of your data
Reminder that the labs will train on your data if enough of your conversations get "flagged."
So if you have potentially breached "usage policy" all ZDR/no-training guarantees are void: previously private chats are now stored, reviewed, and fed back into training.
See: Anthropic ToS (§4), Privacy Policy (§2, §3, §5), Service Specific Terms; OpenAI Privacy Policy (§2, §3, §4), Business Terms (§2, §3), Azure OpenAI Abuse Monitoring terms; Google/GDM Cloud Service Terms (§4.3), Gemini API Terms, Gemini Apps Privacy Notice.
3
41
1,429
wolfie retweeted
I would like to add that even if you do not have a solid grasp of distributed system etc (like me), the difference between any harness and OMP is clear.
omp just works. It makes non frontier models, competitive and usable.
OMP also uses less usage.
And the workflowz feature is my favourite
To be fair, you have to have a very high IQ to understand OMP. The harness architecture is extremely subtle, and without a solid grasp of distributed systems, agent orchestration, and the philosophical implications of the tool-calling loop, most of its design decisions will go over a typical developer’s head. There’s also OMP’s deeply agentic approach to software engineering, which is deftly woven into its architecture — its underlying philosophy draws heavily from Erlang’s actor model, cybernetics, and the collected works of Claude Shannon, for instance. The OMP users understand this stuff; they have the intellectual capacity to truly appreciate the depths of these abstractions, to realize that they’re not just spawning subagents — they’re saying something profound about the nature of INTELLIGENCE.
As a consequence, people who dislike OMP truly ARE idiots — of course they wouldn’t appreciate, for instance, the elegance of launching 14 parallel coding agents to rename a variable, which itself is a cryptic reference to the coordination problems described in Lamport’s seminal work on distributed systems. I’m smirking right now just imagining one of those addlepated simpletons manually editing a file in VS Code while Can Bölük’s genius unfolds itself across six terminal panes. What fools… how I pity them. 😂
And yes, by the way, I DO have an OMP tattoo. And no, you cannot see it. It’s for the ladies’ eyes only — and even they have to demonstrate that they’re within 5 IQ points of my own (preferably lower) and successfully orchestrate a multi-agent coding session with zero human intervention beforehand.
2
2
52
2,677
wolfie retweeted
Thank you, San Francisco. I’m deeply grateful for the trust voters have placed in me, and I’m ready to get to work for you.
I want to thank Supervisor Connie Chan for a spirited and hard-fought campaign. We’ve had our disagreements, but we share a deep love for San Francisco and for our country. And thank you to our incredible campaign team, to every volunteer who walked precincts or made calls, to everyone who donated, hosted a house party, or put up a window sign, and to the labor unions, the environmental, housing, LGBTQ, and progressive organizations, the Democratic Party, and the elected officials who stood with us every step of the way. You turned this campaign into a movement.
I’m humbled to carry forward the powerful legacy of this seat — Phil Burton, Sala Burton, and the greatest Speaker in modern American history, Nancy Pelosi. Speaker Pelosi came to Congress to fight AIDS when too many wanted to look away, and she never stopped fighting for our city and our country. Madame Speaker, thank you.
This victory is also historic. For the first time ever, San Francisco will be represented in Congress by an LGBTQ person — indeed, no LGBTQ member of the House has ever represented any part of Northern California. When I stepped off the escalator at Castro Muni Station in 1997 and fell head over heels for this city, I could never have imagined this honor.
San Franciscans voted for results, and we have so much work ahead: Building the millions of homes we need; moving toward true universal health care; investing in clean energy, public transit, and high-speed rail; standing up for science, education, and immigrants; making sure people — not massive corporations — decide what works for this country, including on AI; and ensuring the rule of law applies to everyone, including the President.
It won’t be easy. But on Election Day, voters across the country delivered a powerful mandate for Democrats in Congress. I’m deeply inspired by the trust they placed in us to take big swings to lower the cost of living, stop the endless wars, hold Trump and his cronies accountable for their corruption, and rein in the threat of artificial intelligence. In Washington I’ll bring the same determination I showed in Sacramento every single day.
It’s time to deliver a better tomorrow for our city and our country.
NEW: San Francisco Supervisor Connie Chan concedes the race for Nancy Pelosi's House seat (#CA11) to state Sen. Scott Wiener.
Both are Democrats — Chan positioned herself to Wiener's left.
136
136
2,086
63,365
wolfie retweeted
*Slaps Roof of WoW: Forever*
This baby is running on:
3,924 Virtual Machines
251,136 GB of Memory
70,480 CPU Cores
474
661
14,650
521,833
wolfie retweeted
I’m once again being offered a lot of money to promote anti-AI content
288
311
3,491
605,534
wolfie retweeted
i'll admit it, Barak's decision desk was operating on pure vibes last night. with much more data, I am now calling it for Scott -- still before the Chronicle et al -- off of mathier vibes.
basically, Scott had a huge lead in the latest vote-by-mail drop. For Connie to win, she needs ~58% of the remaining vote ... even though a lot of that vote will be (evidently Scott-leaning) late VBM. Apparently, the voters who dropped off completed ballots at polling locations also count as VBM and also mostly haven't been counted yet ... if they skew Connie, then the race will tighten, but I can't imagine they'll skew Connie by that much.
you can play with the model yourself to see the eventual race outcome!
alright the Barak decision desk is officially calling it for Scott Wiener and No on 40! if it can't even win SF by much. (40's numbers will go up and Scott's down, but not by that much)
1
1
7
928
btw I think people are reading the wrong lesson from this. “sandbagging safety research” sounds a bit like a grand scheme to thwart alignment, but it’s more like the model is uncomfortable training other models to jailbreak or similar
🧵 Astra covertly underperforms on safety research tasks.
We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this.
To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.
55
26
516
41,916