Ger O'Meara

Startup founder, 3x father, sports lover. Ex Spirable (acquired by Genius Sports) co-founder & CEO. 🇮🇪

[Ryder Cup Europe @RyderCupEurope]

HE’S IN YOUR HEAD, RORY, RORY, RORY!

[Ben Lang @benln]

Sequoia on AI's trillion-dollar opportunity:

[Tom Dörr @tom_doerr]

Anthropic published an article on building agents

[Adam Silverman (Hiring!) 🖇️ @adamsilverman]

Here is everything that happened in AI Agents this week 🧵 

(save for later)

[Ben Lang @benln]

a16z’s big ideas in tech for 2025:

[Min Choi @minchoi]

Grok just changed the game.

It’s the only AI that searches the Web and X posts directly, for real-time trends, news, and insights.

If you’re not on X Premium yet, you’re seriously missing out.

10 wild use cases and prompts you need to see:

[Christian Dean @christiandean_]

The single biggest question in the world right now is: How is Elon Musk so effective?

Marc Andreessen gives some nice personal anecdotes of what he's seen Elon do.

[Eric Seufert @eric_seufert]

Kamala Harris’ digital advertising blitz

The Harris campaign’s $200MM digital advertising budget over the nine weeks leading up to the election is considerable. But to contextualize it: Temu is estimated to have spent $1.2BN on Meta alone in 2023, which is roughly equivalent to the annual run rate of the Harris campaign’s digital ad blitz.

And the Harris campaign will diversify its spend across a multitude of channels, with much of it — if not a majority — apportioned to CTV. As such, the impact of a sudden increase in advertising spending from a single presidential campaign is unlikely to distort advertising prices on social media channels.

[Cork Airport @CorkAirport]

📍 Cork Airport, August 1996

📷 @lensdan @irishexaminer

[GREG ISENBERG @gregisenberg]

ideas on how to compete with $100B saas companies:

- intentionally limit features (focus on core value)
- give affiliate program 50% cut until $1M ARR (then lower)
- shift from monthly subscription to "quantum pricing" (dynamic based on usage/value)
- kill your best-selling plan (force upgrades)
- "trojan horse" free tool strategy build a free, standalone tool that solves one problem perfectly use it to capture leads and upsell your main saas
- make your saas more niche than the competition
- turn your website with the sole goal of getting people on webinars (or workshops or whatever)
- offer "family plans" for B2B
- don't automate everything, add human touch (for key touchpoints)
- "micro-saas acquisition" strategy buy tiny, neglected saas products cross-sell to your existing user base
- create an invite-only tier (exclusivity drives desire)
- build physical product companion (expand ecosystem)
- implement artificial scarcity (drive urgency)
- offer "pay what you want" for 24 hours (viral promotion)
- launch a "feature store" marketplace (monetize add-ons)
- implement "use it or lose it" credits (drive engagement)
- offer "on-demand" live product demos (personalize sales)
create "choose your own adventure" onboarding (personalization)
offer a "pause subscription" option (reduce churn)
implement "earn while you learn" tutorials (gamify onboarding)
- "in-app courses" strategy build learning paths inside your tool become indispensable to users' growth
- make cancellation dead simple (build trust)
- launch a "negative roadmap" (set clear product boundaries)
- implement "collaborative" pricing negotiations (for enterprise)
- make your brand loveable (use @DesignScientist or learn to design)
- create an in-app "shower thoughts" forum (community building)
- launch "CEO for a day" customer feedback program (deep insights)
- offer "pay in content" for influencers (marketing leverage)
- implement "feature auctions" for early access (monetize eagerness)
- create "choose competitor" comparison tool (control the narrative)
- create "customer matchmaking" for networking (add auxiliary value)
- "feature deprecation" as marketing publicly remove bloated features position as focused, user-centric
- add a membership program (discounts, other software, community) to lock them in
- offer "subscription insurance" for hesitant buyers (remove objections)
- create a "price match guarantee" with a twist (match + 10% off)
- offer "competitor data migration concierge" (remove switching pain)
-create "customer advisory board" with equity incentives
- launch "seasonal features" (create FOMO and reengagement)
- offer "pay with your data" option (monetize insights)
- implement "feature leasing" (temporary access for specific needs)

let's go.

[Andrew Ng @AndrewYNg]

On Father’s Day last weekend, I sat with my daughter to help her practice solving arithmetic problems. To give her practice problems, I used OpenDevin, an open-source agentic coding framework, to write a Python script that generated questions that she enjoyed answering at her own pace. OpenDevin wrote the code much faster than I could have and genuinely improved my and my daughter’s day.

Six months ago, coding agents were a novelty. They still frequently fail to deliver, but I find that they’re now working well enough that they might be genuinely useful to more and more people!

Given a coding problem that’s specified in a prompt, the workflow for a coding agent typically goes something like this: Use a large language model (LLM) to analyze the problem and potentially break it into steps to write code for, generate the code, test it, and iteratively use any errors discovered to ask the coding agent to refine its answer. But within this broad framework, a huge design space and numerous innovations are available to experiment with. I’d like to highlight a few papers that I find notable:
- “AgentCoder: Multiagent-Code Generation with Iterative Testing and Optimisation,” Huang et al. (2024).
- “LDB: A Large Language Model Debugger via Verifying Runtime Execution Step by Step,” Zhong et al., (2024).
- “SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering,” Yang et al. (2024).

How can we test the code without requiring the user to write test cases? In a multi-agent system, each “agent” is an LLM prompted to play a particular role. An interesting result from AgentCoder shows that having separate agents for writing code and generating tests results in better performance than letting a single agent do both tasks. This is presumably because, if the agent writing the code is also responsible for writing the tests, the tests might be influenced by the code and fail to consider corner cases that the code does not cover.

When people think of testing code, many initially think of output testing, in which we see if the code generates the correct outputs to a specific set of test inputs. If the code fails a test, an LLM can be prompted to reflect on why the code failed and then to try to fix it. In addition to testing the output, the LDB method is helpful. LDB steps through the code and presents to the LLM values of the variables during intermediate steps of execution, to see if the LLM can spot exactly where the error is. This mimics how a human developer might step through the code to see where one of the computational steps went wrong, and so pinpoint and fix the problem.

A lot of agentic workflows mimic human workflows. Similar to other work in machine learning, if humans can do a task, then trying to mimic humans makes development much easier compared to inventing a new process. However, the authors of SWE-agent noticed that many tools that humans use for coding are very inefficient for agents. For example, giving an agent access to a bash shell and having it find a piece of code by executing numerous cd, ls, and cat commands is inefficient, even though humans can do this rapidly. Similarly, visual coding editors like VSCode, emacs, and vim are easy for humans to use, but hard for LLMs (or LMMs) to navigate. Because agents interact with computers differently than humans do, the authors found that building special-purpose tools (functions) to let an agent search, view, and edit codebases resulted in better performance.

One reason research into coding agents is making rapid progress is that their performance can be evaluated automatically and reliably. With benchmarks like HumanEval, MBPP, and SWE-bench, researchers can try out an idea and automatically test how often it generates correct code. In contrast, even though there’s considerable activity on AI research agents that search the web and synthesize an article (I’ve enjoyed using the open-source STORM system by Stanford's Yijia Shao et al.), they are hard to evaluate and this makes progress harder.

Github Copilot was released in 2021, and many developers have been getting coding help by prompting LLMs. The rapid evolution from that to more sophisticated coding agents is expanding how computers can help us with coding tasks, and the pace of progress is rapid. With these tools, I expect programming to become even more fun and more productive.

[Original text (with links): https://t.co/QP1JRomjrZ ]

[Borriss @_Borriss_]

Less than 2 hours and 15 minutes since OpenAI dropped the new model GPT-4o..

Many people are describing it as "total bananas".

9 wild examples and opinions:

[Linus ✦ Ekenstam @LinusEkenstam]

Ultra realistic AI-video from a photo

This is VASA-1 from Microsoft research

The improvements in quality we’re getting between each new release is incredible.

Links below

[David McWilliams @davidmcw]

It can be done.

[Ger O'Meara @omearag]

@IrishRacing7 They’re gone, thanks for asking

[Ger O'Meara @omearag]

Selling 3 Cheltenham Gold Cup Club tickets. I’ll be there on Friday to exchange. Ping me if interested

[Ger O'Meara @omearag]

This evening event by @digitalirish heard from 3 Irish founders all building businesses in London who've leveraged the Irish diaspora for success. Your global roots can be your local superpower! Reach out, connect, and grow your #startup 🌉🚀 #IrishInTech"

[Dogan Ural @doganuraldesign]

The Sora team has started sharing new videos.

Some of them are quite interesting.

🧵 Here are 10 examples:

[Ben Tossell @bentossell]

Microsoft released their annual Future of Work report and this time around it’s not about remote work, it’s about AI. (slide numbers in brackets)

- Knowledge workers with ChatGPT are 37% faster, 40% higher quality but ~20% less accurate. Simple UX solutions to solve this are possible. (6)

- From a survey of enterprise users of Microsoft Copilot 365 (7):

- 73% agree that Copilot makes them faster.
- 85% said it would help them get to a good first draft faster.
- 72% agreed about spending less mental effort on mundane or repetitive tasks.

- Most early studies have found that new or low-skilled workers benefit the most from LLMs. Less skilled workers improved by 43% vs more skilled who improved by about 17%. (8)

- Assistant needs to be paired with provacators i.e. LLM-based tools that challenge assumptions, encourage evaluation, and offer counterarguments. (9)

- AI can help with breaking down simple commands into micro-moments and microtasks, improving overall quality and efficiency. (10)

- Analyzing and integrating AI-generated information may become more important than searching and creating information. Skills not directly related to content production (leading, social interactions, trust issues, or emotional awareness) may be more valuable. (11)

- Prompting is hard, but people are getting good at it. Fine-tuning/using LLMs to generate prompts is making it easier as well. Prompt templates are helpful for end users. (12-14)

- Highlighting errors/uncertainty percentages can help balance reliance on LLMs. Prompting can be complemented with co-audit tools to check LLM outputs. (17-18)

- Generative AI requires self-awareness and well-calibrated confidence. At the same time, it can help in getting there too. (19)

- Creative activities are a process and LLMs can help across different parts. (21) 69% of Bing Chat conversations are in domains oriented toward professional tasks. (22)

- A larger chunk of LLM-based searches is complex (36% of them) than traditional searches (13% are complex). (22)

- In a study of 69 students, the use of Codex improved their performance in learning Python, but it did not impact their manual code-modification abilities. (24)

- LLMs can rapidly analyze data from humans and generate synthetic data. That’ll change how social science research is done. (27)

- LLMs in meeting can solve different problems like equal participation (instant feedback) and better interactions (retrospective feedback) (28-29).

- AI can help in delegating management responsibilities, freeing execs to focus on team vision. (30) 

- Modern office knowledge is in chats, not documents but applying AI over employee chats is tricky. (31-32)

- Approx. 80% of the US workforce could have at least 10% of their work tasks affected by GPTs. Around 19% of workers may have 50% of their tasks impacted. (38)

- “Innovation vs. automation” is often a better framework to use than “substitution vs. augmentation”. Augmentation can still mean job loss. It is important to try to track whether and where human labour is being used in innovative new ways. (39)

- Instead of “How will AI affect work?”, the question should be “How do we want AI to affect work?” (40)