Richard Nuttall

Interested in AI, Generative AI, AGI, Tesla, SpaceX, Renewables, war in Ukraine, LFTRs

[Richard Nuttall @RichardANuttall]

When I was 27, I built the first UK commercial Internet service - PIPEX. You can achieve a lot at that age, applying energy and drive to really make things happen.

Age provides experience and wisdom, but also slows you down and establishes prejudices.

Coming to a problem fresh allows First Principles Thinking and power through blockers.

[Richard Nuttall @RichardANuttall]

I think we are well past the point of AGI by any reasonable measure. I can have a long design and architecture discussion with Claude Code or Grok Build/Cursor, finalise the Plan and then set it going.
It then builds in an hour what it would have taken a good developer at least a week to build, with a really good completion quality.
There are still blind spots, and improvements to make, but somewhere in the last 3-4 months, we went past AGI.
If we are being picky, there might be a few disciplines where it isn't quite AGI yet, but in a big and growing set of disciplines is is already there.

[Dr. Matthew M. Wielicki @MatthewWielicki]

Some folks always need to complain about something.

[Richard Nuttall @RichardANuttall]

@saeverley @axios Almonds is ~3.6 billion gallons per day, almost all in California.

[Richard Nuttall @RichardANuttall]

Makes a lot of sense. The Superchargers are already connected to good local electricity supply. If they have batteries, or local solar, they will be in a great position to take advantage of off-peak electricity to power the AI compute. Much better use for electricity thatn Bitcoin mining

[Bull Theory @BullTheoryio]

TESLA QUIETLY REVEALED A MASSIVE AI INFRASTRUCTURE PLAY.

Tesla filed a trademark application for "MEGAPOD", signaling plans to turn its Supercharger network into a massive distributed AI computing platform.

The USPTO filing describes MEGAPOD as - "Modular data center hardware systems for artificial intelligence computing, comprised of computer servers, computer hardware for artificial intelligence."

The wording suggests Tesla is developing self-contained, scalable AI server modules that can be deployed across thousands of locations rather than building only centralized data centers.

This will act like a modular unit combining AI compute hardware with Tesla energy systems (Megapack, Powerwall, Superchargers) to act as a single node for power management and AI workloads.

The filing follows Musk's March comments about turning Tesla's 7 GW Supercharger network into one of the world's largest distributed AI compute platforms.

If realized, Tesla's 'MEGAPOD' Could Turn Superchargers Into AI Data Centers.

[Richard Nuttall @RichardANuttall]

@p_ferragu one for the long term. Ely Cathedral near me took 300 years to build and Sagrada Familia is 144 years and counting. Maybe once we have the coming time of abundance when money doesn't matter and we can just tell the Optimi to "Go build me a Cathedral"

[Richard Nuttall @RichardANuttall]

Excellent decision. I have been using Grok Build with Composer 2.5 and the combination rocks. Really looking forward to the next iterations.

[SpaceX @SpaceX]

SpaceX has exercised the option to acquire @cursor_ai in an all-stock transaction with the goal of building the world’s most useful AI models.

For the past few months, SpaceXAI has been jointly training a model with Cursor, which will be released in Cursor and Grok Build soon.

We look forward to working closely with the Cursor team to advance our frontier AI capabilities

[Richard Nuttall @RichardANuttall]

@SpaceX @cursor_ai Excellent decision. I have been using Grok Build with Composer 2.5 and the combination rocks. Really looking forward to the next iterations.

[Richard Nuttall @RichardANuttall]

I think, exactly as I put in my original reply that after the bull rally, investors (institutional and retail) thought "There are upcoming IPOs, including SpaceX and Anthropic and I will take some profit, put that cash aside ready for SpaceX IPO next week".
Given the IPOs was 4X oversubscribed, the retail cash out was ta least $100B".
Those were A contributory factor (not 100%, but also not 0%) to the initial weakness, which then triggered more selloffs.

[Richard Nuttall @RichardANuttall]

https://t.co/9XfKNaDrHt
"Evidence so far is that retail may be saving some dry powder for these upcoming IPOs," Vanda global macro strategist Viraj Patel said. "At this point in the calendar year, we would normally expect slightly stronger activity than what we're currently seeing — and so something seems to be holding retail back."

[Richard Nuttall @RichardANuttall]

SpaceX IPO UK Allocation policy $SPCX :
As per the Allocation Policy for UK retail, investors in the Retail Offer who applied for up to 20 shares have been allocated in full, rounded down to the nearest whole Share.

Allocations were capped at 1,000 shares. This means if you applied for more than 1,000 shares, the maximum you could receive is 1,000. If you applied for more than 20 shares you were allocated 21.13% of your excess application up to the maximum of 1,000 shares.

[Richard Nuttall @RichardANuttall]

@8090_Factory @meet_tie "replaced vendors worth nearly $5B in combined market cap". That has to be the most meaningless statistic ever. My takeaway is that they don't use Grok since then the figure would be $1.7T.

[Richard Nuttall @RichardANuttall]

@alojohhardcore @Sagarforx1 I am not saying it is the only factor, and your analysis is good, but, if e.g. Investors are focused on the IPO and making potential over-subscriptions, that's a lot of cash committed and attention taken from other opportunities.

[Richard Nuttall @RichardANuttall]

@alojohhardcore No mention of SpaceX IPO?
Pulling $78B+ out of the market in prep for the IPO must remove some level of support for current prices and trigger some of the profit taking. I know you don't want to express opinions on SpaceX until after IPO, but it is taking investor attention.

[Richard Nuttall @RichardANuttall]

@JOBhakdi I use Claude Code at work (on Bedrock) and Grok Build at home.
Claude Code is very slow on Bedrock, but currently has better understanding of intent. Grok Build has a better TUI but is behind and improving very fast.

[Richard Nuttall @RichardANuttall]

@skcd42 the /release-notes only shows the latest release note, (each update overwrites the last, it isn't a log) so unless you update frequently, it is easy to miss older updates. Can we see at least the last few updates ?

[Richard Nuttall @RichardANuttall]

From heavy, real-world use over the last couple of weeks on non-trivial projects, here are the areas that still feel clearly early-beta:

1. Plan Mode Review Experience
For anything beyond a trivial task, reviewing the plan requires clicking through multiple screens. The generated plan.md is buried in a hidden session directory and hard to locate.
It would be much better if it offered to save the plan in an easily accessible location (or the project root) so it can be reviewed comfortably in a normal Markdown editor.

2. Default Web UI/UX Quality

When generating web interfaces, the default output still looks weak. The system should have a strong awareness of established design systems and guidelines (Fluent, Material, HIG, etc.) and proactively apply appropriate tokens and patterns instead of producing ad-hoc styling.

3. Over-Amplification of Simple Feedback

Casual requests like “that panel is too big” frequently get interpreted and escalated across iterations into much stronger, permanent constraints than intended. The agent would benefit from better scope retention and explicit confirmation before turning a small tweak into a hard architectural rule.

4. Completion Drive in Long Sessions

In extended implementation runs, the process often stalls part-way through and waits for the user to supply the next prompt rather than driving all the way through to a complete, reviewed, and validated result. Stronger momentum to finish the job would help a lot on larger pieces of work.

5. Transparency on Long Tool-Call Sequences

During longer tasks the agent will sometimes run a long stream of tool calls with almost no commentary on the overall plan or whether the current approach is sub-optimal. A short status note when it suspects it is on a less-than-ideal path would let the user stay oriented and intervene earlier.

6. Release Artifact & Build Discipline

There is still no first-class concept of “this binary is a production artifact that must be built, staged to the right location, and validated.” Sessions often end with a manual “remember to copy it” step. This becomes annoying as the number of native tools grows.

7. Output Verbosity Control

The default personality is still quite verbose and explanatory. Getting genuinely minimal, tool-output-only behaviour for routine work currently requires heavy manual tuning of instructions. Lightweight, first-class controls for output density would be a big improvement.

8. Windows Filesystem & Cross-Account Hygiene

A few Windows-specific problems have already caused real pain: accidental creation of reserved device names (nul) that break Syncthing, and leaving non-ASCII characters in scripts that then produce mojibake on other accounts running classic PowerShell. Built-in guardrails here would prevent recurring breakage.

9. Proactive Debt Detection & Legacy Removal

While the agent is good at cleaning up once something is explicitly called out as technical debt, it rarely surfaces conflicts with established rules on its own (e.g. “this is re-introducing a pattern we moved away from”). It also rarely drives the final aggressive removal of old code once a migration is complete.

10. Long-Session Memory & Coordination

For work that spans many sessions, better automatic capture of key decisions and pruning of superseded context would reduce the amount of manual coordination required. Some concept of “ongoing tracked effort with phases” would also help.

Happy to provide more concrete examples or detail on any of these via DM if useful.

[Richard Nuttall @RichardANuttall]

@DellAnnaLuca Is it too much to ask your LLM to check all the citations before submitting the paper it wrote for you ?

[Richard Nuttall @RichardANuttall]

@somansu_chauhan @grok @xai @elonmusk Confirmed, I get the same. I wanted an update to a recipe in the same thread and Grok 4.3 couldn't see it

[MATLAB @MATLAB]

Introducing the MATLAB Agentic Toolkit

Get started 👉 https://t.co/efiOXdSM6z

[Richard Nuttall @RichardANuttall]

@wholemars 2 points do not make an exponential curve.