Pinned Tweet
Charlie Wells with @Bloomberg reached out to me to ask about my thoughts on VCX and publicly traded funds that hold private AI and tech company shares.
After confirming he was not a Russian bot (always important financial due diligence) I had a great conversation about a trend I start paying attention to:
Public-market investors are looking for ways to gain exposure to private AI and tech companies before they IPO.
My interest in AI started years ago after reading The Singularity Is Near, which completely changed how I thought about technology and society.
The next five years in AI are going to be wild.
Link to the Bloomberg article below
After confirming he was not a Russian bot (always important financial due diligence) I had a great conversation about a trend I start paying attention to:
Public-market investors are looking for ways to gain exposure to private AI and tech companies before they IPO.
My interest in AI started years ago after reading The Singularity Is Near, which completely changed how I thought about technology and society.
The next five years in AI are going to be wild.
Link to the Bloomberg article below
2
0
14
3,030
Claude Opus 4.8 wrote better peer reviews than paid human experts in all 132 decided head-to-head pairs judged by GPT-5.5, in a new Harvard paper by Gary King and Steven Worthington.
• 124 manuscripts and 134 human reviews from Harvard's rapidPeer service, 2019 to 2026; 88% of ed those human reviews high or very high quality
• Three former journal editors, judging 10 blinded pairs, picked the AI review in 26 of 30 judgments, 87%
• 5.5% of AI claims about the manuscripts were judged incorrect vs 15.9% for the humans, with zero fabricated sources
• On a length-matched subset, AI still won 44 of 44
The limits: the editor check covers only 10 pairs, and the study did not measure whether revised papers got better.
• 124 manuscripts and 134 human reviews from Harvard's rapidPeer service, 2019 to 2026; 88% of ed those human reviews high or very high quality
• Three former journal editors, judging 10 blinded pairs, picked the AI review in 26 of 30 judgments, 87%
• 5.5% of AI claims about the manuscripts were judged incorrect vs 15.9% for the humans, with zero fabricated sources
• On a length-matched subset, AI still won 44 of 44
The limits: the editor check covers only 10 pairs, and the study did not measure whether revised papers got better.
0
0
0
80
May give Antigravity another shot if you can use Sol and Astra, I am over the UI of Codex, probably a mistake by OpenAI to allow this
openai.com/index/openai-m…
openai.com/index/openai-m…
0
0
2
151
Does anyone else feel like Roger Goodell is a complete imposter who is a highly overpaid caricature of an executive?…asking for a friend
nbcsports.com/nfl/profootbal…
nbcsports.com/nfl/profootbal…
0
0
1
57
RelativelySmart retweeted
I am building Accordant: AI Voice agents that use approved business knowledge, follow clear rules, and hand off to a person when needed.
The site and practical guide are live. Take a look and let me know what you think.
ai-explanation.com
The site and practical guide are live. Take a look and let me know what you think.
ai-explanation.com
3
1
2
59
Select Opus 5.5 and prompt it with "Do you know Voxel Bench? Do not search" to detect if Anthropic is routing to a newer "superior" model.
Earlier this month, several users reported that Opus 5.5 is routing to another, more advanced model that generates significantly better output.
Read more 🗞️
testingcatalog.com/anthropic-test…
1
0
1
282
RelativelySmart retweeted
most people are not psychologically ready for the possibility of a permanent underclass
315
159
2,514
270,945
Anthropic is putting 5% of its shares into a new Anthropic Public Benefit Fund, subject to stockholder approval, with a stake that can grow as the company's valuation rises.
> Escalator: up to 2% more for every 50% rise in company value above a baseline Anthropic has not disclosed, at the board's discretion
> Structure: a wholly owned LLC rather than a 501(c)(3), so it can shift between nonprofit donations and direct distributions to specific populations
> Limits: none of the fund's equity can be used for lobbying or political giving, in the US or abroad
> Total: with the cofounders' 80% equity pledge and an employee match used by about half of staff, roughly 20% of shares are now committed
> First named use: AI's impact on the labor market, where Anthropic has already committed $200 million
Governance and the valuation milestones are due "in the coming months."
> Escalator: up to 2% more for every 50% rise in company value above a baseline Anthropic has not disclosed, at the board's discretion
> Structure: a wholly owned LLC rather than a 501(c)(3), so it can shift between nonprofit donations and direct distributions to specific populations
> Limits: none of the fund's equity can be used for lobbying or political giving, in the US or abroad
> Total: with the cofounders' 80% equity pledge and an employee match used by about half of staff, roughly 20% of shares are now committed
> First named use: AI's impact on the labor market, where Anthropic has already committed $200 million
Governance and the valuation milestones are due "in the coming months."
1
0
2
270
Define Public Good and how exactly that will work?
We’re committing a portion of Anthropic’s shares to the Anthropic Public Benefit Fund, a new wholly owned entity dedicated entirely to the public benefit.
It will start off with 5% of our shares. For every 50% increase in the company’s value over a set baseline, we can contribute up to an extra 2%, subject to our board’s discretion.
In total, this means that roughly 20% of Anthropic’s shares are committed to the public good. Our co-founders have already pledged to donate 80% of their equity, and around half our employees are donating a portion of their equity as part of a donation matching program we’ve run since our founding.
In the post below, we share more details about our new fund.
anthropic.com/news/public-be…
1
0
2
472
Microsoft has dropped its deal to anchor Nscale's Loughton, Essex site, which the UK government called the "largest UK sovereign AI datacentre" at its 2025 announcement. That leaves the site with no publicly announced customer.
- Plan: 50MW scalable to 90MW, about 23,000 Nvidia osoft Azure from Q1 2027
- Power: UK Power Networks reportedly told Nscale the grid cannot supply enough until the early to mid 2030s
- Ofgem counted 315 datacentres waiting to connect in June, 73GW of demand against a 45GW national peak
- Exposure: Microsoft and Anthropic make up about 85% of Nscale's contracts, per its IPO filing
- Timing: Nscale is seeking up to $3B in a US IPO at a potential $35B valuation
Nscale says it is exploring alternative power for Loughton and that the London wide grid issue has no impact on its other contracts. Microsoft is understood to remain on its Norway project.
- Plan: 50MW scalable to 90MW, about 23,000 Nvidia osoft Azure from Q1 2027
- Power: UK Power Networks reportedly told Nscale the grid cannot supply enough until the early to mid 2030s
- Ofgem counted 315 datacentres waiting to connect in June, 73GW of demand against a 45GW national peak
- Exposure: Microsoft and Anthropic make up about 85% of Nscale's contracts, per its IPO filing
- Timing: Nscale is seeking up to $3B in a US IPO at a potential $35B valuation
Nscale says it is exploring alternative power for Loughton and that the London wide grid issue has no impact on its other contracts. Microsoft is understood to remain on its Norway project.
2
1
2
148
RelativelySmart retweeted
what I've learned from this post is that almost no one is AGI-pilled and that they don't take seriously the possibility of tens to hundreds of millions of AIs working 24/7 at 10–100x human speed
Surveyed Anthropic staff's answers imply 17.1% GDP growth in 2030
70
37
924
65,815
China's commerce and science ministries proposed export controls on 23 AI technologies on Oct 16, including the weights of frontier models.
Weights: checkpoints, source code, and architecture are restricted at 10^26 training FLOPs, 2 trillion parameters, or a global top-10 thresholds include 20 or more sub-tasks per instruction, 60 or more concurrent agents, or 200 or more integrated tools
Compute: cluster techniques above 10,000 AI accelerators
Exempt: anything already open-sourced or published, plus serving a model abroad by API, app, or SaaS
Of the 23 AI items, 6 would be prohibited and 17 restricted. Comments close Nov 15, and pre-export registration took effect on publication day.
Weights: checkpoints, source code, and architecture are restricted at 10^26 training FLOPs, 2 trillion parameters, or a global top-10 thresholds include 20 or more sub-tasks per instruction, 60 or more concurrent agents, or 200 or more integrated tools
Compute: cluster techniques above 10,000 AI accelerators
Exempt: anything already open-sourced or published, plus serving a model abroad by API, app, or SaaS
Of the 23 AI items, 6 would be prohibited and 17 restricted. Comments close Nov 15, and pre-export registration took effect on publication day.
1
1
5
1,227
See a lot of complaining on here every week about OpenAI plans but I have been using upwards of ~300--400M tokens a day the last 90 days in Codex. While yes @thsottiaux gives out occasional resets when I am close to 0%. *I have utilized 4-5 banked resets but I rarely if ever run t using Astra xHigh for more than 1 or 2 threads. This is on a $100/month Premium Business Seat btw which I upgraded to ~2 months ago.
0
0
1
553
Meta released the weights for Muse Spark Open on Oct 15: 816B total parameters, 42B active per token, a 1M-token context, and text, image, video, and audio input.
SPEC: 256 experts per MoE layer, 8 routed per token; BF16 checkpoint about 1.5 TB, plus NVFP4
DESK: runs on two 512 GB M5 Ultra Mac Studios linked over Thunderbolt 5 with MLX
SCORES: MCP-Atlas 89.2, OSWorld Verified 80.4, Terminal-Bench 2.1 76.4, AA Intelligence Index 39
LICENSE: free of charge with a "Built with Muse Spark Open" label; model-as-a-service firms above $20M revenue need a separate Meta deal
Hugging Face measured 18 tokens/s for a 2-bit quant on a single 512 GB M3 Ultra.
SPEC: 256 experts per MoE layer, 8 routed per token; BF16 checkpoint about 1.5 TB, plus NVFP4
DESK: runs on two 512 GB M5 Ultra Mac Studios linked over Thunderbolt 5 with MLX
SCORES: MCP-Atlas 89.2, OSWorld Verified 80.4, Terminal-Bench 2.1 76.4, AA Intelligence Index 39
LICENSE: free of charge with a "Built with Muse Spark Open" label; model-as-a-service firms above $20M revenue need a separate Meta deal
Hugging Face measured 18 tokens/s for a 2-bit quant on a single 512 GB M3 Ultra.
0
0
2
190
Anthropic's monorepo merges rose from about 2,500 per day in May to over 30,000 in September. A seven-engineer team built Scion, an in-house version control system, and finished the Git migration in five months.
• Main-branch history alone is almost 100 GB; compressed working copy over 25 GB
• About 25 million CI checkouts per day before counting developer and agent sandboxes
• Scion footprint is 200x smaller than prior Git mirrors in CPU and memory
• Checkouts cut from minutes to seconds; CI alone freed 25% CPU capacity
• Claude Mythos Preview rewrote an ingestion pipeline to 20x less memory and 4x faster
Anthropic says Scion is not a general Git replacement and will partner with GitHub, GitLab, and OSS on agent-scale server designs.
• Main-branch history alone is almost 100 GB; compressed working copy over 25 GB
• About 25 million CI checkouts per day before counting developer and agent sandboxes
• Scion footprint is 200x smaller than prior Git mirrors in CPU and memory
• Checkouts cut from minutes to seconds; CI alone freed 25% CPU capacity
• Claude Mythos Preview rewrote an ingestion pipeline to 20x less memory and 4x faster
Anthropic says Scion is not a general Git replacement and will partner with GitHub, GitLab, and OSS on agent-scale server designs.
1
0
3
113
Always seems to be news on Quantum chip breakthroughs but their biggest problem is error rate, a real 40x reduction in error rate by Google’s Willow processor team is likely a notable & newsworthy achievement.
researchers demonstrated something called magic state cultivation on their Willow processor, reducing errors by 40x and achieving 99.99% logical fidelity.
For those who don’t know, producing high-quality magic states is one of the hardest and biggest bottlenecks in building computers.
Some estimates say that up to 95% of a quantum computer’s resources could be dedicated to generating them, and somehow Google managed to experimentally demonstrate a way to purify those states using “coordinated error-detection measurements” and transfer them between different quantum error-correcting codes.
They still need to scale this up, but if they do, this could significantly reduce the hardware required to build useful quantum computers.
1
0
1
177